Category Archives: Research Blog

The latest news on Google Research

A decoder-only foundation model for time-series forecasting

Posted by Rajat Sen and Yichen Zhou, Google Research

Time-series forecasting is ubiquitous in various domains, such as retail, finance, manufacturing, healthcare and natural sciences. In retail use cases, for example, it has been observed that improving demand forecasting accuracy can meaningfully reduce inventory costs and increase revenue. Deep learning (DL) models have emerged as a popular approach for forecasting rich, multivariate, time-series data because they have proven to perform well in a variety of settings (e.g., DL models dominated the M5 competition leaderboard).

At the same time, there has been rapid progress in large foundation language models used for natural language processing (NLP) tasks, such as translation, retrieval-augmented generation, and code completion. These models are trained on massive amounts of textual data derived from a variety of sources like common crawl and open-source code that allows them to identify patterns in languages. This makes them very powerful zero-shot tools; for instance, when paired with retrieval, they can answer questions about and summarize current events.

Despite DL-based forecasters largely outperforming traditional methods and progress being made in reducing training and inference costs, they face challenges: most DL architectures require long and involved training and validation cycles before a customer can test the model on a new time-series. A foundation model for time-series forecasting, in contrast, can provide decent out-of-the-box forecasts on unseen time-series data with no additional training, enabling users to focus on refining forecasts for the actual downstream task like retail demand planning.

To that end, in “A decoder-only foundation model for time-series forecasting”, we introduce TimesFM, a single forecasting model pre-trained on a large time-series corpus of 100 billion real world time-points. Compared to the latest large language models (LLMs), TimesFM is much smaller (200M parameters), yet we show that even at such scales, its zero-shot performance on a variety of unseen datasets of different domains and temporal granularities come close to the state-of-the-art supervised approaches trained explicitly on these datasets. Later this year we plan to make this model available for external customers in Google Cloud Vertex AI.


A decoder-only foundation model for time-series forecasting

LLMs are usually trained in a decoder-only fashion that involves three steps. First, text is broken down into subwords called tokens. Then, the tokens are fed into stacked causal transformer layers that produce an output corresponding to each input token (it cannot attend to future tokens). Finally, the output corresponding to the i-th token summarizes all the information from previous tokens and predicts the (i+1)-th token. During inference, the LLM generates the output one token at a time. For example, when prompted with “What is the capital of France?”, it might generate the token “The”, then condition on “What is the capital of France? The” to generate the next token “capital” and so on until it generates the complete answer: “The capital of France is Paris”.

A foundation model for time-series forecasting should adapt to variable context (what we observe) and horizon (what we query the model to forecast) lengths, while having enough capacity to encode all patterns from a large pretraining dataset. Similar to LLMs, we use stacked transformer layers (self-attention and feedforward layers) as the main building blocks for the TimesFM model. In the context of time-series forecasting, we treat a patch (a group of contiguous time-points) as a token that was popularized by a recent long-horizon forecasting work. The task then is to forecast the (i+1)-th patch of time-points given the i-th output at the end of the stacked transformer layers.

However, there are several key differences from language models. Firstly, we need a multilayer perceptron block with residual connections to convert a patch of time-series into a token that can be input to the transformer layers along with positional encodings (PE). For that, we use a residual block similar to our prior work in long-horizon forecasting. Secondly, at the other end, an output token from the stacked transformer can be used to predict a longer length of subsequent time-points than the input patch length, i.e., the output patch length can be larger than the input patch length.

Consider a time-series of length 512 time-points being used to train a TimesFM model with input patch length 32 and output patch length 128. During training, the model is simultaneously trained to use the first 32 time-points to forecast the next 128 time-points, the first 64 time-points to forecast time-points 65 to 192, the first 96 time-points to forecast time-points 97 to 224 and so on. During inference, suppose the model is given a new time-series of length 256 and tasked with forecasting the next 256 time-points into the future. The model will first generate the future predictions for time-points 257 to 384, then condition on the initial 256 length input plus the generated output to generate time-points 385 to 512. On the other hand, if in our model the output patch length was equal to the input patch length of 32 then for the same task we would have to go through eight generation steps instead of just the two above. This increases the chances of more errors accumulating and therefore, in practice, we see that a longer output patch length yields better performance for long-horizon forecasting

TimesFM architecture.


Pretraining data

Just like LLMs get better with more tokens, TimesFM requires a large volume of legitimate time series data to learn and improve. We have spent a great amount of time creating and assessing our training datasets, and the following is what we have found works best:

Synthetic data helps with the basics. Meaningful synthetic time-series data can be generated using statistical models or physical simulations. These basic temporal patterns can teach the model the grammar of time series forecasting.

Real-world data adds real-world flavor. We comb through available public time series datasets, and selectively put together a large corpus of 100 billion time-points. Among these datasets there are Google Trends and Wikipedia Pageviews, which track what people are interested in, and that nicely mirrors trends and patterns in many other real-world time series. This helps TimesFM understand the bigger picture and generalize better when provided with domain-specific contexts not seen during training.


Zero-shot evaluation results

We evaluate TimesFM zero-shot on data not seen during training using popular time-series benchmarks. We observe that TimesFM performs better than most statistical methods like ARIMA, ETS and can match or outperform powerful DL models like DeepAR, PatchTST that have been explicitly trained on the target time-series.

We used the Monash Forecasting Archive to evaluate TimesFM’s out-of-the-box performance. This archive contains tens of thousands of time-series from various domains like traffic, weather, and demand forecasting covering frequencies ranging from few minutes to yearly data. Following existing literature, we inspect the mean absolute error (MAE) appropriately scaled so that it can be averaged across the datasets. We see that zero-shot (ZS) TimesFM is better than most supervised approaches, including recent deep learning models. We also compare TimesFM to GPT-3.5 for forecasting using a specific prompting technique proposed by llmtime(ZS). We demonstrate that TimesFM performs better than llmtime(ZS) despite being orders of magnitude smaller.

Scaled MAE (the lower the better) of TimesFM(ZS) against other supervised and zero-shot approaches on Monash datasets.

Most of the Monash datasets are short or medium horizon, i.e., the prediction length is not too long. We also test TimesFM on popular benchmarks for long horizon forecasting against a recent state-of-the-art baseline PatchTST (and other long-horizon forecasting baselines). In the next figure, we plot the MAE on ETT datasets for the task of predicting 96 and 192 time-points into the future. The metric has been calculated on the last test window of each dataset (as done by the llmtime paper). We see that TimesFM not only surpasses the performance of llmtime(ZS) but also matches that of the supervised PatchTST model explicitly trained on the respective datasets.

Last window MAE (the lower the better) of TimesFM(ZS) against llmtime(ZS) and long-horizon forecasting baselines on ETT datasets.


Conclusion

We train a decoder-only foundation model for time-series forecasting using a large pretraining corpus of 100B real world time-points, the majority of which was search interest time-series data derived from Google Trends and pageviews from Wikipedia. We show that even a relatively small 200M parameter pretrained model that uses our TimesFM architecture displays impressive zero-shot performance on a variety of public benchmarks from different domains and granularities.


Acknowledgements

This work is the result of a collaboration between several individuals across Google Research and Google Cloud, including (in alphabetical order): Abhimanyu Das, Weihao Kong, Andrew Leach, Mike Lawrence, Alex Martin, Rajat Sen, Yang Yang and Yichen Zhou.

Source: Google AI Blog


Intervening on early readouts for mitigating spurious features and simplicity bias

Posted by Rishabh Tiwari, Pre-doctoral Researcher, and Pradeep Shenoy, Research Scientist, Google Research

Machine learning models in the real world are often trained on limited data that may contain unintended statistical biases. For example, in the CELEBA celebrity image dataset, a disproportionate number of female celebrities have blond hair, leading to classifiers incorrectly predicting “blond” as the hair color for most female faces — here, gender is a spurious feature for predicting hair color. Such unfair biases could have significant consequences in critical applications such as medical diagnosis.

Surprisingly, recent work has also discovered an inherent tendency of deep networks to amplify such statistical biases, through the so-called simplicity bias of deep learning. This bias is the tendency of deep networks to identify weakly predictive features early in the training, and continue to anchor on these features, failing to identify more complex and potentially more accurate features.

With the above in mind, we propose simple and effective fixes to this dual challenge of spurious features and simplicity bias by applying early readouts and feature forgetting. First, in “Using Early Readouts to Mediate Featural Bias in Distillation”, we show that making predictions from early layers of a deep network (referred to as “early readouts”) can automatically signal issues with the quality of the learned representations. In particular, these predictions are more often wrong, and more confidently wrong, when the network is relying on spurious features. We use this erroneous confidence to improve outcomes in model distillation, a setting where a larger “teacher” model guides the training of a smaller “student” model. Then in “Overcoming Simplicity Bias in Deep Networks using a Feature Sieve”, we intervene directly on these indicator signals by making the network “forget” the problematic features and consequently look for better, more predictive features. This substantially improves the model’s ability to generalize to unseen domains compared to previous approaches. Our AI Principles and our Responsible AI practices guide how we research and develop these advanced applications and help us address the challenges posed by statistical biases.

Animation comparing hypothetical responses from two models trained with and without the feature sieve.

Early readouts for debiasing distillation

We first illustrate the diagnostic value of early readouts and their application in debiased distillation, i.e., making sure that the student model inherits the teacher model’s resilience to feature bias through distillation. We start with a standard distillation framework where the student is trained with a mixture of label matching (minimizing the cross-entropy loss between student outputs and the ground-truth labels) and teacher matching (minimizing the KL divergence loss between student and teacher outputs for any given input).

Suppose one trains a linear decoder, i.e., a small auxiliary neural network named as Aux, on top of an intermediate representation of the student model. We refer to the output of this linear decoder as an early readout of the network representation. Our finding is that early readouts make more errors on instances that contain spurious features, and further, the confidence on those errors is higher than the confidence associated with other errors. This suggests that confidence on errors from early readouts is a fairly strong, automated indicator of the model’s dependence on potentially spurious features.

Illustrating the usage of early readouts (i.e., output from the auxiliary layer) in debiasing distillation. Instances that are confidently mispredicted in the early readouts are upweighted in the distillation loss.

We used this signal to modulate the contribution of the teacher in the distillation loss on a per-instance basis, and found significant improvements in the trained student model as a result.

We evaluated our approach on standard benchmark datasets known to contain spurious correlations (Waterbirds, CelebA, CivilComments, MNLI). Each of these datasets contain groupings of data that share an attribute potentially correlated with the label in a spurious manner. As an example, the CelebA dataset mentioned above includes groups such as {blond male, blond female, non-blond male, non-blond female}, with models typically performing the worst on the {non-blond female} group when predicting hair color. Thus, a measure of model performance is its worst group accuracy, i.e., the lowest accuracy among all known groups present in the dataset. We improved the worst group accuracy of student models on all datasets; moreover, we also improved overall accuracy in three of the four datasets, showing that our improvement on any one group does not come at the expense of accuracy on other groups. More details are available in our paper.

Comparison of Worst Group Accuracies of different distillation techniques relative to that of the Teacher model. Our method outperforms other methods on all datasets.

Overcoming simplicity bias with a feature sieve

In a second, closely related project, we intervene directly on the information provided by early readouts, to improve feature learning and generalization. The workflow alternates between identifying problematic features and erasing identified features from the network. Our primary hypothesis is that early features are more prone to simplicity bias, and that by erasing (“sieving”) these features, we allow richer feature representations to be learned.

Training workflow with feature sieve. We alternate between identifying problematic features (using training iteration) and erasing them from the network (using forgetting iteration).

We describe the identification and erasure steps in more detail:

  • Identifying simple features: We train the primary model and the readout model (AUX above) in conventional fashion via forward- and back-propagation. Note that feedback from the auxiliary layer does not back-propagate to the main network. This is to force the auxiliary layer to learn from already-available features rather than create or reinforce them in the main network.
  • Applying the feature sieve: We aim to erase the identified features in the early layers of the neural network with the use of a novel forgetting loss, Lf , which is simply the cross-entropy between the readout and a uniform distribution over labels. Essentially, all information that leads to nontrivial readouts are erased from the primary network. In this step, the auxiliary network and upper layers of the main network are kept unchanged.

We can control specifically how the feature sieve is applied to a given dataset through a small number of configuration parameters. By changing the position and complexity of the auxiliary network, we control the complexity of the identified- and erased features. By modifying the mixing of learning and forgetting steps, we control the degree to which the model is challenged to learn more complex features. These choices, which are dataset-dependent, are made via hyperparameter search to maximize validation accuracy, a standard measure of generalization. Since we include “no-forgetting” (i.e., the baseline model) in the search space, we expect to find settings that are at least as good as the baseline.

Below we show features learned by the baseline model (middle row) and our model (bottom row) on two benchmark datasets — biased activity recognition (BAR) and animal categorization (NICO). Feature importance was estimated using post-hoc gradient-based importance scoring (GRAD-CAM), with the orange-red end of the spectrum indicating high importance, while green-blue indicates low importance. Shown below, our trained models focus on the primary object of interest, whereas the baseline model tends to focus on background features that are simpler and spuriously correlated with the label.

Feature importance scoring using GRAD-CAM on activity recognition (BAR) and animal categorization (NICO) generalization benchmarks. Our approach (last row) focuses on the relevant objects in the image, whereas the baseline (ERM; middle row) relies on background features that are spuriously correlated with the label.

Through this ability to learn better, generalizable features, we show substantial gains over a range of relevant baselines on real-world spurious feature benchmark datasets: BAR, CelebA Hair, NICO and ImagenetA, by margins up to 11% (see figure below). More details are available in our paper.

Our feature sieve method improves accuracy by significant margins relative to the nearest baseline for a range of feature generalization benchmark datasets.

Conclusion

We hope that our work on early readouts and their use in feature sieving for generalization will both spur the development of a new class of adversarial feature learning approaches and help improve the generalization capability and robustness of deep learning systems.


Acknowledgements

The work on applying early readouts to debiasing distillation was conducted in collaboration with our academic partners Durga Sivasubramanian, Anmol Reddy and Prof. Ganesh Ramakrishnan at IIT Bombay. We extend our sincere gratitude to Praneeth Netrapalli and Anshul Nasery for their feedback and recommendations. We are also grateful to Nishant Jain, Shreyas Havaldar, Rachit Bansal, Kartikeya Badola, Amandeep Kaur and the whole cohort of pre-doctoral researchers at Google Research India for taking part in research discussions. Special thanks to Tom Small for creating the animation used in this post.

Source: Google AI Blog


MobileDiffusion: Rapid text-to-image generation on-device

Posted by Yang Zhao, Senior Software Engineer, and Tingbo Hou, Senior Staff Software Engineer, Core ML

Text-to-image diffusion models have shown exceptional capabilities in generating high-quality images from text prompts. However, leading models feature billions of parameters and are consequently expensive to run, requiring powerful desktops or servers (e.g., Stable Diffusion, DALL·E, and Imagen). While recent advancements in inference solutions on Android via MediaPipe and iOS via Core ML have been made in the past year, rapid (sub-second) text-to-image generation on mobile devices has remained out of reach.

To that end, in “MobileDiffusion: Subsecond Text-to-Image Generation on Mobile Devices”, we introduce a novel approach with the potential for rapid text-to-image generation on-device. MobileDiffusion is an efficient latent diffusion model specifically designed for mobile devices. We also adopt DiffusionGAN to achieve one-step sampling during inference, which fine-tunes a pre-trained diffusion model while leveraging a GAN to model the denoising step. We have tested MobileDiffusion on iOS and Android premium devices, and it can run in half a second to generate a 512x512 high-quality image. Its comparably small model size of just 520M parameters makes it uniquely suited for mobile deployment.

      
Rapid text-to-image generation on-device.

Background

The relative inefficiency of text-to-image diffusion models arises from two primary challenges. First, the inherent design of diffusion models requires iterative denoising to generate images, necessitating multiple evaluations of the model. Second, the complexity of the network architecture in text-to-image diffusion models involves a substantial number of parameters, regularly reaching into the billions and resulting in computationally expensive evaluations. As a result, despite the potential benefits of deploying generative models on mobile devices, such as enhancing user experience and addressing emerging privacy concerns, it remains relatively unexplored within the current literature.

The optimization of inference efficiency in text-to-image diffusion models has been an active research area. Previous studies predominantly concentrate on addressing the first challenge, seeking to reduce the number of function evaluations (NFEs). Leveraging advanced numerical solvers (e.g., DPM) or distillation techniques (e.g., progressive distillation, consistency distillation), the number of necessary sampling steps have significantly reduced from several hundreds to single digits. Some recent techniques, like DiffusionGAN and Adversarial Diffusion Distillation, even reduce to a single necessary step.

However, on mobile devices, even a small number of evaluation steps can be slow due to the complexity of model architecture. Thus far, the architectural efficiency of text-to-image diffusion models has received comparatively less attention. A handful of earlier works briefly touches upon this matter, involving the removal of redundant neural network blocks (e.g., SnapFusion). However, these efforts lack a comprehensive analysis of each component within the model architecture, thereby falling short of providing a holistic guide for designing highly efficient architectures.


MobileDiffusion

Effectively overcoming the challenges imposed by the limited computational power of mobile devices requires an in-depth and holistic exploration of the model's architectural efficiency. In pursuit of this objective, our research undertakes a detailed examination of each constituent and computational operation within Stable Diffusion’s UNet architecture. We present a comprehensive guide for crafting highly efficient text-to-image diffusion models culminating in the MobileDiffusion.

The design of MobileDiffusion follows that of latent diffusion models. It contains three components: a text encoder, a diffusion UNet, and an image decoder. For the text encoder, we use CLIP-ViT/L14, which is a small model (125M parameters) suitable for mobile. We then turn our focus to the diffusion UNet and image decoder.


Diffusion UNet

As illustrated in the figure below, diffusion UNets commonly interleave transformer blocks and convolution blocks. We conduct a comprehensive investigation of these two fundamental building blocks. Throughout the study, we control the training pipeline (e.g., data, optimizer) to study the effects of different architectures.

In classic text-to-image diffusion models, a transformer block consists of a self-attention layer (SA) for modeling long-range dependencies among visual features, a cross-attention layer (CA) to capture interactions between text conditioning and visual features, and a feed-forward layer (FF) to post-process the output of attention layers. These transformer blocks hold a pivotal role in text-to-image diffusion models, serving as the primary components responsible for text comprehension. However, they also pose a significant efficiency challenge, given the computational expense of the attention operation, which is quadratic to the sequence length. We follow the idea of UViT architecture, which places more transformer blocks at the bottleneck of the UNet. This design choice is motivated by the fact that the attention computation is less resource-intensive at the bottleneck due to its lower dimensionality.

Our UNet architecture incorporates more transformers in the middle, and skips self-attention (SA) layers at higher resolutions.

Convolution blocks, in particular ResNet blocks, are deployed at each level of the UNet. While these blocks are instrumental for feature extraction and information flow, the associated computational costs, especially at high-resolution levels, can be substantial. One proven approach in this context is separable convolution. We observed that replacing regular convolution layers with lightweight separable convolution layers in the deeper segments of the UNet yields similar performance.

In the figure below, we compare the UNets of several diffusion models. Our MobileDiffusion exhibits superior efficiency in terms of FLOPs (floating-point operations) and number of parameters.

Comparison of some diffusion UNets.

Image decoder

In addition to the UNet, we also optimized the image decoder. We trained a variational autoencoder (VAE) to encode an RGB image to an 8-channel latent variable, with 8× smaller spatial size of the image. A latent variable can be decoded to an image and gets 8× larger in size. To further enhance efficiency, we design a lightweight decoder architecture by pruning the original’s width and depth. The resulting lightweight decoder leads to a significant performance boost, with nearly 50% latency improvement and better quality. For more details, please refer to our paper.

VAE reconstruction. Our VAE decoders have better visual quality than SD (Stable Diffusion).

Decoder   #Params (M)     PSNR↑     SSIM↑     LPIPS↓  
SD 49.5 26.7 0.76 0.037
Ours 39.3 30.0 0.83 0.032
Ours-Lite     9.8 30.2 0.84 0.032

Quality evaluation of VAE decoders. Our lite decoder is much smaller than SD, with better quality metrics, including peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS).

One-step sampling

In addition to optimizing the model architecture, we adopt a DiffusionGAN hybrid to achieve one-step sampling. Training DiffusionGAN hybrid models for text-to-image generation encounters several intricacies. Notably, the discriminator, a classifier distinguishing real data and generated data, must make judgments based on both texture and semantics. Moreover, the cost of training text-to-image models can be extremely high, particularly in the case of GAN-based models, where the discriminator introduces additional parameters. Purely GAN-based text-to-image models (e.g., StyleGAN-T, GigaGAN) confront similar complexities, resulting in highly intricate and expensive training.

To overcome these challenges, we use a pre-trained diffusion UNet to initialize the generator and discriminator. This design enables seamless initialization with the pre-trained diffusion model. We postulate that the internal features within the diffusion model contain rich information of the intricate interplay between textual and visual data. This initialization strategy significantly streamlines the training.

The figure below illustrates the training procedure. After initialization, a noisy image is sent to the generator for one-step diffusion. The result is evaluated against ground truth with a reconstruction loss, similar to diffusion model training. We then add noise to the output and send it to the discriminator, whose result is evaluated with a GAN loss, effectively adopting the GAN to model a denoising step. By using pre-trained weights to initialize the generator and the discriminator, the training becomes a fine-tuning process, which converges in less than 10K iterations.

Illustration of DiffusionGAN fine-tuning.

Results

Below we show example images generated by our MobileDiffusion with DiffusionGAN one-step sampling. With such a compact model (520M parameters in total), MobileDiffusion can generate high-quality diverse images for various domains.

Images generated by our MobileDiffusion

We measured the performance of our MobileDiffusion on both iOS and Android devices, using different runtime optimizers. The latency numbers are reported below. We see that MobileDiffusion is very efficient and can run within half a second to generate a 512x512 image. This lightning speed potentially enables many interesting use cases on mobile devices.

Latency measurements (s) on mobile devices.

Conclusion

With superior efficiency in terms of latency and size, MobileDiffusion has the potential to be a very friendly option for mobile deployments given its capability to enable a rapid image generation experience while typing text prompts. And we will ensure any application of this technology will be in-line with Google’s responsible AI practices.


Acknowledgments

We like to thank our collaborators and contributors that helped bring MobileDiffusion to on-device: Zhisheng Xiao, Yanwu Xu, Jiuqiang Tang, Haolin Jia, Lutz Justen, Daniel Fenner, Ronald Wotzlaw, Jianing Wei, Raman Sarokin, Juhyun Lee, Andrei Kulik, Chuo-Ling Chang, and Matthias Grundmann.

Source: Google AI Blog


Mixed-input matrix multiplication performance optimizations

Posted by Manish Gupta, Staff Software Engineer, Google Research

AI-driven technologies are weaving themselves into the fabric of our daily routines, with the potential to enhance our access to knowledge and boost our overall productivity. The backbone of these applications lies in large language models (LLMs). LLMs are memory-intensive and typically require specialized hardware accelerators to efficiently deliver tens of exaflops of computing power. This blog post shows how we can start addressing the computational challenges by utilizing memory more effectively.

The bulk of an LLM’s memory and compute are consumed by weights in matrix multiplication operations. Using narrower data types reduces memory consumption. For example, storing weights in the 8-bit integer (i.e., U8 or S8) data type reduces the memory footprint by 4× relative to single-precision (F32) and 2× relative to half-precision (F16) or bfloat16 (BF16). Furthermore, previous work has shown that LLM models running matrix multiplications with weights in S8 and input in F16 (preserving higher precision of the user-input) is an effective method for increasing the efficiency with acceptable trade-offs in accuracy. This technique is known as weight-only quantization and requires efficient implementation of matrix multiplication with mixed-inputs, e.g., half-precision input multiplied with 8-bits integer. Hardware accelerators, including GPUs, support a fixed set of data types, and thus, mixed-input matrix multiplication requires software transformations to map to the hardware operations.

To that end, in this blog we focus on mapping mixed-input matrix multiplication onto the NVIDIA Ampere architecture. We present software techniques addressing data type conversion and layout conformance to map mixed-input matrix multiplication efficiently onto hardware-supported data types and layouts. Our results show that the overhead of additional work in software is minimal and enables performance close to the peak hardware capabilities. The software techniques described here are released in the open-source NVIDIA/CUTLASS repository.

Memory footprint for an 175B parameter LLM model with various data types formats.

The matrix-multiply-accumulate operation

Modern AI hardware accelerators such as Google’s TPU and NVIDIA’s GPU multiply matrices natively in the hardware by targeting Tensor Cores, which are specialized processing elements to accelerate matrix operations, particularly for AI workloads. In this blog, we focus on NVIDIA Ampere Tensor Cores, which provide the matrix-multiply-accumulate (mma) operation. For the rest of the blog the reference to mma is for Ampere Tensor Cores. The supported data types, shapes, and data layout of the two input matrices (called operands) for the mma operation are fixed in hardware. This means that matrix multiplications with various data types and larger shapes are implemented in the software by tiling the problem onto hardware-supported data types, shapes, and layouts.

The Tensor Core mma operation is defined by specifying two input matrices (e.g., A & B, shown below) to produce a result matrix, C. The mma operation natively supports mixed-precision. Mixed-precision Tensor Cores allow mixing input (A and B) data type with the result (C) data type. In contrast, mixed-input matrix multiplication involves mixing the input data types, and it is not supported by the hardware, so it needs to be implemented in the software.

Tensor Core operation of M-by-N-by-K on input matrix A of M-by-K and matrix B of K-by-N produces output matrix C of M-by-N.

Challenges of mixed-input matrix multiplication

To simplify the discussion, we restrict to a specific example of mixed-input matrix multiplication: F16 for user input and U8 for the model weights (written as F16 * U8). The techniques described here work for various combinations of mixed-input data types.

A GPU programmer can access a hierarchy of memory, including global memory, shared memory, and registers, which are arranged in order of decreasing capacity but increasing speed. NVIDIA Ampere Tensor Core mma operations consume input matrices from registers. Furthermore, input and output matrices are required to conform to a layout of data within a group of 32 threads known as a warp. The supported data type and layout within a warp are fixed for an mma operation, so to implement mixed-input multiplication efficiently, it is necessary to solve the challenges of data type conversion and layout conformance in software.


Data type conversion

The mma operation requires two input matrices with the same data type. Thus, mixed-input matrix multiplication, where one of the operands is stored in U8 in global memory and other in F16, requires a data type conversion from U8 to F16. The conversion will bring two operands to F16, mapping the mixed-input matrix multiplication to hardware-supported mixed-precision Tensor Cores. Given the large number of weights, there are a large number of such operations, and our techniques show how to reduce their latency and improve performance.


Layout conformance

The mma operation also requires the layout of two input matrices, within the registers of a warp, to be conformat with hardware specification. The layout for the input matrix B of U8 data type in mixed-input matrix multiplication (F16 * U8) needs to conform with the converted F16 data type. This is called layout conformance and needs to be achieved in the software.

The figure below shows an mma operation consuming matrix A and matrix B from registers to produce matrix C in registers, distributed across one warp. The thread T0 is highlighted and zoomed in to show the weight matrix B goes through data type conversion and needs a layout conformance to be able to map to the hardware-supported Tensor Core operation.

The mapping of mixed-input (F32 = F16 * U8) operation in software to natively supported warp-level Tensor Cores in hardware (F32 = F16 * F16). (Original figure source Developing CUDA kernels to push Tensor Cores to the Absolute Limit on NVIDIA A100.)

Software strategies addressing challenges

A typical data type conversion involves a sequence of operations on 32-bit registers, shown below. Each rectangular block represents a register and the adjoining text are the operations. The entire sequence shows the conversion from 4xU8 to 2x(2xF16). The sequence involves roughly 10 operations.

NumericArrayConvertor from 4xU8 to 2x(2xF16) in 32-bit registers.

There are many ways of achieving layout conformance. Two of the existing solutions are:

  1. Narrower bitwidth shared memory loads: In this approach, threads issue narrow bitwidth memory loads moving the U8 data from shared memory to registers. This results in two 32-bit registers, with each register containing 2xF16 values (shown above for the matrix B’s thread T0). The narrower shared memory load achieves layout conformance directly into registers without needing any shuffles; however, it does not utilize the full shared memory bandwidth.
  2. Pre-processing in global memory: An alternative strategy involves rearranging the data within the global memory (one level above the shared memory in memory hierarchy), allowing wider shared memory loads. This approach maximizes the shared memory bandwidth utilization and ensures that the data is loaded in a conformant layout directly in the registers. Although the rearrangement process can be executed offline prior to the LLM deployment, ensuring no impact on the application performance, it introduces an additional, non-trivial hardware-specific pre-processing step that requires an extra program to rearrange the data. NVIDIA/FasterTransformer adopts this method to effectively address layout conformance challenges.

Optimized software strategies

To further optimize and reduce the overhead of data type conversion and layout conformance, we have implemented FastNumericArrayConvertor and FragmentShuffler, respectively.

FastNumericArrayConvertor operates on 4xU8 in 32-bit registers without unpacking individual 1xU8 values. Furthermore, it uses less expensive arithmetic operations which reduces the number of instructions and increases the speed of the conversion.

The conversion sequence for U8-to-F16 is shown below. The operations use packed 32b registers, avoiding explicit unpacking and packing. FastNumericArrayConvertor uses the permute byte to rearrange bytes of 4xU8 into two registers. Additionally, FastNumericArrayConvertor does not use expensive integer to floating-point conversion instructions and employs vectorized operations to obtain the packed results in two 32-bit registers containing 2x(2xF16) values. The FastNumericArrayConvertor for U8-to-F16 approximately uses six operations, a 1.6× reduction relative to the approach shown above.

FastNumericArrayConvertor utilizes permute bytes and packed arithmetic, reducing the number of instructions in the data type conversion.

FragmentShuffler handles the layout conformance by shuffling data in a way that allows the use of wider bitwidth load operation, increasing shared memory bandwidth utilization and reducing the total number of operations.

NVIDIA Ampere architecture provides a load matrix instruction (ldmatrix). The ldmatrix is a warp-level operation, where 32 threads of a warp move the data from shared memory to registers in the shape and layout that mma matrix A and B consume. The use of ldmatrix reduces the number of load instructions and increases the memory bandwidth utilization. Since the ldmatrix instruction moves U8 data to registers, the layout after the load conforms with U8*U8 mma operation, and not with F16*F16 mma operation. We implemented FragmentShuffler to rearrange the data within registers using shuffle (shfl.sync) operations to achieve the layout conformance.

The most significant contribution of this work is to achieve layout conformance through register shuffles, avoiding offline pre-processing in global memory or narrower bitwidth shared memory loads. Furthermore, we provide implementations for FastNumericArrayConvertor covering data type conversion from U8-to-F16, S8-to-F16, U8-to-BF16, and S8-to-BF16.


Performance results

We measured the performance of eight mixed-input variants of our method (shown below in blue and red; varying the data types of matrix A and B) and two mixed-precision data types (shown in green) on an NVIDIA A100 SXM chip. The performance results are shown in FLOPS (higher is better). Notably, the first eight matrix-multipications require additional operations relative to the last two, because the mixed-precision variants directly target hardware-accelerated Tensor Core operations and do not need data type conversion and layout conformance. Even so, our approach demonstrates mixed-input matrix multiplication performance only slightly below or on par with mixed-precision.

Mixed-input matrix multiplication performance on NVIDIA A100 40GB SMX4 chip for a compute-bound matrix problem shape m=3456, n=4096, k=2048.

Acknowledgements

We would like to mention several folks who have contributed through technical brainstorming and improving the blog post including, Quentin Colombet, Jacques Pienaar, Allie Culp, Calin Cascaval, Ashish Gondimalla, Matt Walsh, Marek Kolodziej, and Aman Bhatia. We would like to thank our NVIDIA partners Rawn Henry, Pradeep Ramani, Vijay Thakkar, Haicheng Wu, Andrew Kerr, Matthew Nicely, and Vartika Singh.

Source: Google AI Blog


Exphormer: Scaling transformers for graph-structured data

Posted by Ameya Velingker, Research Scientist, Google Research, and Balaji Venkatachalam, Software Engineer, Google

Graphs, in which objects and their relations are represented as nodes (or vertices) and edges (or links) between pairs of nodes, are ubiquitous in computing and machine learning (ML). For example, social networks, road networks, and molecular structure and interactions are all domains in which underlying datasets have a natural graph structure. ML can be used to learn the properties of nodes, edges, or entire graphs.

A common approach to learning on graphs are graph neural networks (GNNs), which operate on graph data by applying an optimizable transformation on node, edge, and global attributes. The most typical class of GNNs operates via a message-passing framework, whereby each layer aggregates the representation of a node with those of its immediate neighbors.

Recently, graph transformer models have emerged as a popular alternative to message-passing GNNs. These models build on the success of Transformer architectures in natural language processing (NLP), adapting them to graph-structured data. The attention mechanism in graph transformers can be modeled by an interaction graph, in which edges represent pairs of nodes that attend to each other. Unlike message passing architectures, graph transformers have an interaction graph that is separate from the input graph. The typical interaction graph is a complete graph, which signifies a full attention mechanism that models direct interactions between all pairs of nodes. However, this creates quadratic computational and memory bottlenecks that limit the applicability of graph transformers to datasets on small graphs with at most a few thousand nodes. Making graph transformers scalable has been considered one of the most important research directions in the field (see the first open problem here).

A natural remedy is to use a sparse interaction graph with fewer edges. Many sparse and efficient transformers have been proposed to eliminate the quadratic bottleneck for sequences, however, they do not generally extend to graphs in a principled manner.

In “Exphormer: Sparse Transformers for Graphs”, presented at ICML 2023, we address the scalability challenge by introducing a sparse attention framework for transformers that is designed specifically for graph data. The Exphormer framework makes use of expander graphs, a powerful tool from spectral graph theory, and is able to achieve strong empirical results on a wide variety of datasets. Our implementation of Exphormer is now available on GitHub.


Expander graphs

A key idea at the heart of Exphormer is the use of expander graphs, which are sparse yet well-connected graphs that have some useful properties — 1) the matrix representation of the graphs have similar linear-algebraic properties as a complete graph, and 2) they exhibit rapid mixing of random walks, i.e., a small number of steps in a random walk from any starting node is enough to ensure convergence to a “stable” distribution on the nodes of the graph. Expanders have found applications to diverse areas, such as algorithms, pseudorandomness, complexity theory, and error-correcting codes.

A common class of expander graphs are d-regular expanders, in which there are d edges from every node (i.e., every node has degree d). The quality of an expander graph is measured by its spectral gap, an algebraic property of its adjacency matrix (a matrix representation of the graph in which rows and columns are indexed by nodes and entries indicate whether pairs of nodes are connected by an edge). Those that maximize the spectral gap are known as Ramanujan graphs — they achieve a gap of d - 2*√(d-1), which is essentially the best possible among d-regular graphs. A number of deterministic and randomized constructions of Ramanujan graphs have been proposed over the years for various values of d. We use a randomized expander construction of Friedman, which produces near-Ramanujan graphs.

Expander graphs are at the heart of Exphormer. A good expander is sparse yet exhibits rapid mixing of random walks, making its global connectivity suitable for an interaction graph in a graph transformer model.

Exphormer replaces the dense, fully-connected interaction graph of a standard Transformer with edges of a sparse d-regular expander graph. Intuitively, the spectral approximation and mixing properties of an expander graph allow distant nodes to communicate with each other after one stacks multiple attention layers in a graph transformer architecture, even though the nodes may not attend to each other directly. Furthermore, by ensuring that d is constant (independent of the size of the number of nodes), we obtain a linear number of edges in the resulting interaction graph.


Exphormer: Constructing a sparse interaction graph

Exphormer combines expander edges with the input graph and virtual nodes. More specifically, the sparse attention mechanism of Exphormer builds an interaction graph consisting of three types of edges:

  • Edges from the input graph (local attention)
  • Edges from a constant-degree expander graph (expander attention)
  • Edges from every node to a small set of virtual nodes (global attention)
Exphormer builds an interaction graph by combining three types of edges. The resulting graph has good connectivity properties and retains the inductive bias of the input dataset graph while still remaining sparse.

Each component serves a specific purpose: the edges from the input graph retain the inductive bias from the input graph structure (which typically gets lost in a fully-connected attention module). Meanwhile, expander edges allow good global connectivity and random walk mixing properties (which spectrally approximate the complete graph with far fewer edges). Finally, virtual nodes serve as global “memory sinks” that can directly communicate with every node. While this results in additional edges from each virtual node equal to the number of nodes in the input graph, the resulting graph is still sparse. The degree of the expander graph and the number of virtual nodes are hyperparameters to tune for improving the quality metrics.

Furthermore, since we use an expander graph of constant degree and a small constant number of virtual nodes for the global attention, the resulting sparse attention mechanism is linear in the size of the original input graph, i.e., it models a number of direct interactions on the order of the total number of nodes and edges.

We additionally show that Exphormer is as expressive as the dense transformer and obeys universal approximation properties. In particular, when the sparse attention graph of Exphormer is augmented with self loops (edges connecting a node to itself), it can universally approximate continuous functions [1, 2].


Relation to sparse Transformers for sequences

It is interesting to compare Exphormer to sparse attention methods for sequences. Perhaps the architecture most conceptually similar to our approach is BigBird, which builds an interaction graph by combining different components. BigBird also uses virtual nodes, but, unlike Exphormer, it uses window attention and random attention from an Erdős-Rényi random graph model for the remaining components.

Window attention in BigBird looks at the tokens surrounding a token in a sequence — the local neighborhood attention in Exphormer can be viewed as a generalization of window attention to graphs.

The Erdős-Rényi graph on n nodes, G(n, p), which connects every pair of nodes independently with probability p, also functions as an expander graph for suitably high p. However, a superlinear number of edges (Ω(n log n)) is needed to ensure that an Erdős-Rényi graph is connected, let alone a good expander. On the other hand, the expanders used in Exphormer have only a linear number of edges.


Experimental results

Earlier works have shown the use of full graph Transformer-based models on datasets with graphs of size up to 5,000 nodes. To evaluate the performance of Exphormer, we build upon the celebrated GraphGPS framework [3], which combines both message passing and graph transformers and achieves state-of-the-art performance on a number of datasets. We show that replacing dense attention with Exphormer for the graph attention component in the GraphGPS framework allows one to achieve models with comparable or better performance, often with fewer trainable parameters.

Furthermore, Exphormer notably allows graph transformer architectures to scale well beyond the usual graph size limits mentioned above. Exphormer can scale up to datasets of 10,000+ node graphs, such as the Coauthor dataset, and even beyond to larger graphs such as the well-known ogbn-arxiv dataset, a citation network, which consists of 170K nodes and 1.1 million edges.

Results comparing Exphormer to standard GraphGPS on the five Long Range Graph Benchmark datasets. We note that Exphormer achieved state-of-the-art results on four of the five datasets (PascalVOC-SP, COCO-SP, Peptides-Struct, PCQM-Contact) at the time of the paper’s publication.

Finally, we observe that Exphormer, which creates an overlay graph of small diameter via expanders, exhibits the ability to effectively learn long-range dependencies. The Long Range Graph Benchmark is a suite of five graph learning datasets designed to measure the ability of models to capture long-range interactions. Results show that Exphormer-based models outperform standard GraphGPS models (which were previously state-of-the-art on four out of five datasets at the time of publication).


Conclusion

Graph transformers have emerged as an important architecture for ML that adapts the highly successful sequence-based transformers used in NLP to graph-structured data. Scalability has, however, proven to be a major challenge in enabling the use of graph transformers on datasets with large graphs. In this post, we have presented Exphormer, a sparse attention framework that uses expander graphs to improve scalability of graph transformers. Exphormer is shown to have important theoretical properties and exhibit strong empirical performance, particularly on datasets where it is crucial to learn long range dependencies. For more information, we point the reader to a short presentation video from ICML 2023.


Acknowledgements

We thank our research collaborators Hamed Shirzad and Danica J. Sutherland from The University of British Columbia as well as Ali Kemal Sinop from Google Research. Special thanks to Tom Small for creating the animation used in this post.

Source: Google AI Blog


Introducing ASPIRE for selective prediction in LLMs

Posted by Jiefeng Chen, Student Researcher, and Jinsung Yoon, Research Scientist, Cloud AI Team

In the fast-evolving landscape of artificial intelligence, large language models (LLMs) have revolutionized the way we interact with machines, pushing the boundaries of natural language understanding and generation to unprecedented heights. Yet, the leap into high-stakes decision-making applications remains a chasm too wide, primarily due to the inherent uncertainty of model predictions. Traditional LLMs generate responses recursively, yet they lack an intrinsic mechanism to assign a confidence score to these responses. Although one can derive a confidence score by summing up the probabilities of individual tokens in the sequence, traditional approaches typically fall short in reliably distinguishing between correct and incorrect answers. But what if LLMs could gauge their own confidence and only make predictions when they're sure?

Selective prediction aims to do this by enabling LLMs to output an answer along with a selection score, which indicates the probability that the answer is correct. With selective prediction, one can better understand the reliability of LLMs deployed in a variety of applications. Prior research, such as semantic uncertainty and self-evaluation, has attempted to enable selective prediction in LLMs. A typical approach is to use heuristic prompts like “Is the proposed answer True or False?” to trigger self-evaluation in LLMs. However, this approach may not work well on challenging question answering (QA) tasks.

The OPT-2.7B model incorrectly answers a question from the TriviaQA dataset: “Which vitamin helps regulate blood clotting?” with “Vitamin C”. Without selective prediction, LLMs may output the wrong answer which, in this case, could lead users to take the wrong vitamin. With selective prediction, LLMs will output an answer along with a selection score. If the selection score is low (0.1), LLMs will further output “I don’t know!” to warn users not to trust it or verify it using other sources.

In "Adaptation with Self-Evaluation to Improve Selective Prediction in LLMs", presented at Findings of EMNLP 2023, we introduce ASPIRE — a novel framework meticulously designed to enhance the selective prediction capabilities of LLMs. ASPIRE fine-tunes LLMs on QA tasks via parameter-efficient fine-tuning, and trains them to evaluate whether their generated answers are correct. ASPIRE allows LLMs to output an answer along with a confidence score for that answer. Our experimental results demonstrate that ASPIRE significantly outperforms state-of-the-art selective prediction methods on a variety of QA datasets, such as the CoQA benchmark.


The mechanics of ASPIRE

Imagine teaching an LLM to not only answer questions but also evaluate those answers — akin to a student verifying their answers in the back of the textbook. That's the essence of ASPIRE, which involves three stages: (1) task-specific tuning, (2) answer sampling, and (3) self-evaluation learning.

Task-specific tuning: ASPIRE performs task-specific tuning to train adaptable parameters (θp) while freezing the LLM. Given a training dataset for a generative task, it fine-tunes the pre-trained LLM to improve its prediction performance. Towards this end, parameter-efficient tuning techniques (e.g., soft prompt tuning and LoRA) might be employed to adapt the pre-trained LLM on the task, given their effectiveness in obtaining strong generalization with small amounts of target task data. Specifically, the LLM parameters (θ) are frozen and adaptable parameters (θp) are added for fine-tuning. Only θp are updated to minimize the standard LLM training loss (e.g., cross-entropy). Such fine-tuning can improve selective prediction performance because it not only improves the prediction accuracy, but also enhances the likelihood of correct output sequences.

Answer sampling: After task-specific tuning, ASPIRE uses the LLM with the learned θp to generate different answers for each training question and create a dataset for self-evaluation learning. We aim to generate output sequences that have a high likelihood. We use beam search as the decoding algorithm to generate high-likelihood output sequences and the Rouge-L metric to determine if the generated output sequence is correct.

Self-evaluation learning: After sampling high-likelihood outputs for each query, ASPIRE adds adaptable parameters (θs) and only fine-tunes θs for learning self-evaluation. Since the output sequence generation only depends on θ and θp, freezing θ and the learned θp can avoid changing the prediction behaviors of the LLM when learning self-evaluation. We optimize θs such that the adapted LLM can distinguish between correct and incorrect answers on their own.

The three stages of the ASPIRE framework.

In the proposed framework, θp and θs can be trained using any parameter-efficient tuning approach. In this work, we use soft prompt tuning, a simple yet effective mechanism for learning “soft prompts” to condition frozen language models to perform specific downstream tasks more effectively than traditional discrete text prompts. The driving force behind this approach lies in the recognition that if we can develop prompts that effectively stimulate self-evaluation, it should be possible to discover these prompts through soft prompt tuning in conjunction with targeted training objectives.

Implementation of the ASPIRE framework via soft prompt tuning. We first generate the answer to the question with the first soft prompt and then compute the learned self-evaluation score with the second soft prompt.

After training θp and θs, we obtain the prediction for the query via beam search decoding. We then define a selection score that combines the likelihood of the generated answer with the learned self-evaluation score (i.e., the likelihood of the prediction being correct for the query) to make selective predictions.


Results

To demonstrate ASPIRE’s efficacy, we evaluate it across three question-answering datasets — CoQA, TriviaQA, and SQuAD — using various open pre-trained transformer (OPT) models. By training θp with soft prompt tuning, we observed a substantial hike in the LLMs' accuracy. For example, the OPT-2.7B model adapted with ASPIRE demonstrated improved performance over the larger, pre-trained OPT-30B model using the CoQA and SQuAD datasets. These results suggest that with suitable adaptations, smaller LLMs might have the capability to match or potentially surpass the accuracy of larger models in some scenarios.

When delving into the computation of selection scores with fixed model predictions, ASPIRE received a higher AUROC score (the probability that a randomly chosen correct output sequence has a higher selection score than a randomly chosen incorrect output sequence) than baseline methods across all datasets. For example, on the CoQA benchmark, ASPIRE improves the AUROC from 51.3% to 80.3% compared to the baselines.

An intriguing pattern emerged from the TriviaQA dataset evaluations. While the pre-trained OPT-30B model demonstrated higher baseline accuracy, its performance in selective prediction did not improve significantly when traditional self-evaluation methods — Self-eval and P(True) — were applied. In contrast, the smaller OPT-2.7B model, when enhanced with ASPIRE, outperformed in this aspect. This discrepancy underscores a vital insight: larger LLMs utilizing conventional self-evaluation techniques may not be as effective in selective prediction as smaller, ASPIRE-enhanced models.

Our experimental journey with ASPIRE underscores a pivotal shift in the landscape of LLMs: The capacity of a language model is not the be-all and end-all of its performance. Instead, the effectiveness of models can be drastically improved through strategic adaptations, allowing for more precise, confident predictions even in smaller models. As a result, ASPIRE stands as a testament to the potential of LLMs that can judiciously ascertain their own certainty and decisively outperform larger counterparts in selective prediction tasks.


Conclusion

In conclusion, ASPIRE is not just another framework; it's a vision of a future where LLMs can be trusted partners in decision-making. By honing the selective prediction performance, we're inching closer to realizing the full potential of AI in critical applications.

Our research has opened new doors, and we invite the community to build upon this foundation. We're excited to see how ASPIRE will inspire the next generation of LLMs and beyond. To learn more about our findings, we encourage you to read our paper and join us in this thrilling journey towards creating a more reliable and self-aware AI.


Acknowledgments

We gratefully acknowledge the contributions of Sayna Ebrahimi, Sercan O Arik, Tomas Pfister, and Somesh Jha.

Source: Google AI Blog


AMIE: A research AI system for diagnostic medical reasoning and conversations

Posted by Alan Karthikesalingam and Vivek Natarajan, Research Leads, Google Research

The physician-patient conversation is a cornerstone of medicine, in which skilled and intentional communication drives diagnosis, management, empathy and trust. AI systems capable of such diagnostic dialogues could increase availability, accessibility, quality and consistency of care by being useful conversational partners to clinicians and patients alike. But approximating clinicians’ considerable expertise is a significant challenge.

Recent progress in large language models (LLMs) outside the medical domain has shown that they can plan, reason, and use relevant context to hold rich conversations. However, there are many aspects of good diagnostic dialogue that are unique to the medical domain. An effective clinician takes a complete “clinical history” and asks intelligent questions that help to derive a differential diagnosis. They wield considerable skill to foster an effective relationship, provide information clearly, make joint and informed decisions with the patient, respond empathically to their emotions, and support them in the next steps of care. While LLMs can accurately perform tasks such as medical summarization or answering medical questions, there has been little work specifically aimed towards developing these kinds of conversational diagnostic capabilities.

Inspired by this challenge, we developed Articulate Medical Intelligence Explorer (AMIE), a research AI system based on a LLM and optimized for diagnostic reasoning and conversations. We trained and evaluated AMIE along many dimensions that reflect quality in real-world clinical consultations from the perspective of both clinicians and patients. To scale AMIE across a multitude of disease conditions, specialties and scenarios, we developed a novel self-play based simulated diagnostic dialogue environment with automated feedback mechanisms to enrich and accelerate its learning process. We also introduced an inference time chain-of-reasoning strategy to improve AMIE’s diagnostic accuracy and conversation quality. Finally, we tested AMIE prospectively in real examples of multi-turn dialogue by simulating consultations with trained actors.

AMIE was optimized for diagnostic conversations, asking questions that help to reduce its uncertainty and improve diagnostic accuracy, while also balancing this with other requirements of effective clinical communication, such as empathy, fostering a relationship, and providing information clearly.

Evaluation of conversational diagnostic AI

Besides developing and optimizing AI systems themselves for diagnostic conversations, how to assess such systems is also an open question. Inspired by accepted tools used to measure consultation quality and clinical communication skills in real-world settings, we constructed a pilot evaluation rubric to assess diagnostic conversations along axes pertaining to history-taking, diagnostic accuracy, clinical management, clinical communication skills, relationship fostering and empathy.

We then designed a randomized, double-blind crossover study of text-based consultations with validated patient actors interacting either with board-certified primary care physicians (PCPs) or the AI system optimized for diagnostic dialogue. We set up our consultations in the style of an objective structured clinical examination (OSCE), a practical assessment commonly used in the real world to examine clinicians’ skills and competencies in a standardized and objective way. In a typical OSCE, clinicians might rotate through multiple stations, each simulating a real-life clinical scenario where they perform tasks such as conducting a consultation with a standardized patient actor (trained carefully to emulate a patient with a particular condition). Consultations were performed using a synchronous text-chat tool, mimicking the interface familiar to most consumers using LLMs today.

AMIE is a research AI system based on LLMs for diagnostic reasoning and dialogue.

AMIE: an LLM-based conversational diagnostic research AI system

We trained AMIE on real-world datasets comprising medical reasoning, medical summarization and real-world clinical conversations.

It is feasible to train LLMs using real-world dialogues developed by passively collecting and transcribing in-person clinical visits, however, two substantial challenges limit their effectiveness in training LLMs for medical conversations. First, existing real-world data often fails to capture the vast range of medical conditions and scenarios, hindering the scalability and comprehensiveness. Second, the data derived from real-world dialogue transcripts tends to be noisy, containing ambiguous language (including slang, jargon, humor and sarcasm), interruptions, ungrammatical utterances, and implicit references.

To address these limitations, we designed a self-play based simulated learning environment with automated feedback mechanisms for diagnostic medical dialogue in a virtual care setting, enabling us to scale AMIE’s knowledge and capabilities across many medical conditions and contexts. We used this environment to iteratively fine-tune AMIE with an evolving set of simulated dialogues in addition to the static corpus of real-world data described.

This process consisted of two self-play loops: (1) an “inner” self-play loop, where AMIE leveraged in-context critic feedback to refine its behavior on simulated conversations with an AI patient simulator; and (2) an “outer” self-play loop where the set of refined simulated dialogues were incorporated into subsequent fine-tuning iterations. The resulting new version of AMIE could then participate in the inner loop again, creating a virtuous continuous learning cycle.

Further, we also employed an inference time chain-of-reasoning strategy which enabled AMIE to progressively refine its response conditioned on the current conversation to arrive at an informed and grounded reply.

AMIE uses a novel self-play based simulated dialogue learning environment to improve the quality of diagnostic dialogue across a multitude of disease conditions, specialities and patient contexts.

We tested performance in consultations with simulated patients (played by trained actors), compared to those performed by 20 real PCPs using the randomized approach described above. AMIE and PCPs were assessed from the perspectives of both specialist attending physicians and our simulated patients in a randomized, blinded crossover study that included 149 case scenarios from OSCE providers in Canada, the UK and India in a diverse range of specialties and diseases.

Notably, our study was not designed to emulate either traditional in-person OSCE evaluations or the ways clinicians usually use text, email, chat or telemedicine. Instead, our experiment mirrored the most common way consumers interact with LLMs today, a potentially scalable and familiar mechanism for AI systems to engage in remote diagnostic dialogue.

Overview of the randomized study design to perform a virtual remote OSCE with simulated patients via online multi-turn synchronous text chat.

Performance of AMIE

In this setting, we observed that AMIE performed simulated diagnostic conversations at least as well as PCPs when both were evaluated along multiple clinically-meaningful axes of consultation quality. AMIE had greater diagnostic accuracy and superior performance for 28 of 32 axes from the perspective of specialist physicians, and 24 of 26 axes from the perspective of patient actors.

AMIE outperformed PCPs on multiple evaluation axes for diagnostic dialogue in our evaluations.
Specialist-rated top-k diagnostic accuracy. AMIE and PCPs top-k differential diagnosis (DDx) accuracy are compared across 149 scenarios with respect to the ground truth diagnosis (a) and all diagnoses listed within the accepted differential diagnoses (b). Bootstrapping (n=10,000) confirms all top-k differences between AMIE and PCP DDx accuracy are significant with p <0.05 after false discovery rate (FDR) correction.
Diagnostic conversation and reasoning qualities as assessed by specialist physicians. On 28 out of 32 axes, AMIE outperformed PCPs while being comparable on the rest.

Limitations

Our research has several limitations and should be interpreted with appropriate caution. Firstly, our evaluation technique likely underestimates the real-world value of human conversations, as the clinicians in our study were limited to an unfamiliar text-chat interface, which permits large-scale LLM–patient interactions but is not representative of usual clinical practice. Secondly, any research of this type must be seen as only a first exploratory step on a long journey. Transitioning from a LLM research prototype that we evaluated in this study to a safe and robust tool that could be used by people and those who provide care for them will require significant additional research. There are many important limitations to be addressed, including experimental performance under real-world constraints and dedicated exploration of such important topics as health equity and fairness, privacy, robustness, and many more, to ensure the safety and reliability of the technology.


AMIE as an aid to clinicians

In a recently released preprint, we evaluated the ability of an earlier iteration of the AMIE system to generate a DDx alone or as an aid to clinicians. Twenty (20) generalist clinicians evaluated 303 challenging, real-world medical cases sourced from the New England Journal of Medicine (NEJM) ClinicoPathologic Conferences (CPCs). Each case report was read by two clinicians randomized to one of two assistive conditions: either assistance from search engines and standard medical resources, or AMIE assistance in addition to these tools. All clinicians provided a baseline, unassisted DDx prior to using the respective assistive tools.

Assisted randomized reader study setup to investigate the assistive effect of AMIE to clinicians in solving complex diagnostic case challenges from the New England Journal of Medicine.

AMIE exhibited standalone performance that exceeded that of unassisted clinicians (top-10 accuracy 59.1% vs. 33.6%, p= 0.04). Comparing the two assisted study arms, the top-10 accuracy was higher for clinicians assisted by AMIE, compared to clinicians without AMIE assistance (24.6%, p<0.01) and clinicians with search (5.45%, p=0.02). Further, clinicians assisted by AMIE arrived at more comprehensive differential lists than those without AMIE assistance.

In addition to strong standalone performance, using the AMIE system led to significant assistive effect and improvements in diagnostic accuracy of the clinicians in solving these complex case challenges.

It's worth noting that NEJM CPCs are not representative of everyday clinical practice. They are unusual case reports in only a few hundred individuals so offer limited scope for probing important issues like equity or fairness.


Bold and responsible research in healthcare — the art of the possible

Access to clinical expertise remains scarce around the world. While AI has shown great promise in specific clinical applications, engagement in the dynamic, conversational diagnostic journeys of clinical practice requires many capabilities not yet demonstrated by AI systems. Doctors wield not only knowledge and skill but a dedication to myriad principles, including safety and quality, communication, partnership and teamwork, trust, and professionalism. Realizing these attributes in AI systems is an inspiring challenge that should be approached responsibly and with care. AMIE is our exploration of the “art of the possible”, a research-only system for safely exploring a vision of the future where AI systems might be better aligned with attributes of the skilled clinicians entrusted with our care. It is early experimental-only work, not a product, and has several limitations that we believe merit rigorous and extensive further scientific studies in order to envision a future in which conversational, empathic and diagnostic AI systems might become safe, helpful and accessible.


Acknowledgements

The research described here is joint work across many teams at Google Research and Google Deepmind. We are grateful to all our co-authors - Tao Tu, Mike Schaekermann, Anil Palepu, Daniel McDuff, Jake Sunshine, Khaled Saab, Jan Freyberg, Ryutaro Tanno, Amy Wang, Brenna Li, Mohamed Amin, Sara Mahdavi, Karan Sighal, Shekoofeh Azizi, Nenad Tomasev, Yun Liu, Yong Cheng, Le Hou, Albert Webson, Jake Garrison, Yash Sharma, Anupam Pathak, Sushant Prakash, Philip Mansfield, Shwetak Patel, Bradley Green, Ewa Dominowska, Renee Wong, Juraj Gottweis, Dale Webster, Katherine Chou, Christopher Semturs, Joelle Barral, Greg Corrado and Yossi Matias. We also thank Sami Lachgar, Lauren Winer and John Guilyard for their support with narratives and the visuals. Finally, we are grateful to Michael Howell, James Maynika, Jeff Dean, Karen DeSalvo, Zoubin Gharahmani and Demis Hassabis for their support during the course of this project.


Source: Google AI Blog


AMIE: A research AI system for diagnostic medical reasoning and conversations

Posted by Alan Karthikesalingam and Vivek Natarajan, Research Leads, Google Research

The physician-patient conversation is a cornerstone of medicine, in which skilled and intentional communication drives diagnosis, management, empathy and trust. AI systems capable of such diagnostic dialogues could increase availability, accessibility, quality and consistency of care by being useful conversational partners to clinicians and patients alike. But approximating clinicians’ considerable expertise is a significant challenge.

Recent progress in large language models (LLMs) outside the medical domain has shown that they can plan, reason, and use relevant context to hold rich conversations. However, there are many aspects of good diagnostic dialogue that are unique to the medical domain. An effective clinician takes a complete “clinical history” and asks intelligent questions that help to derive a differential diagnosis. They wield considerable skill to foster an effective relationship, provide information clearly, make joint and informed decisions with the patient, respond empathically to their emotions, and support them in the next steps of care. While LLMs can accurately perform tasks such as medical summarization or answering medical questions, there has been little work specifically aimed towards developing these kinds of conversational diagnostic capabilities.

Inspired by this challenge, we developed Articulate Medical Intelligence Explorer (AMIE), a research AI system based on a LLM and optimized for diagnostic reasoning and conversations. We trained and evaluated AMIE along many dimensions that reflect quality in real-world clinical consultations from the perspective of both clinicians and patients. To scale AMIE across a multitude of disease conditions, specialties and scenarios, we developed a novel self-play based simulated diagnostic dialogue environment with automated feedback mechanisms to enrich and accelerate its learning process. We also introduced an inference time chain-of-reasoning strategy to improve AMIE’s diagnostic accuracy and conversation quality. Finally, we tested AMIE prospectively in real examples of multi-turn dialogue by simulating consultations with trained actors.

AMIE was optimized for diagnostic conversations, asking questions that help to reduce its uncertainty and improve diagnostic accuracy, while also balancing this with other requirements of effective clinical communication, such as empathy, fostering a relationship, and providing information clearly.

Evaluation of conversational diagnostic AI

Besides developing and optimizing AI systems themselves for diagnostic conversations, how to assess such systems is also an open question. Inspired by accepted tools used to measure consultation quality and clinical communication skills in real-world settings, we constructed a pilot evaluation rubric to assess diagnostic conversations along axes pertaining to history-taking, diagnostic accuracy, clinical management, clinical communication skills, relationship fostering and empathy.

We then designed a randomized, double-blind crossover study of text-based consultations with validated patient actors interacting either with board-certified primary care physicians (PCPs) or the AI system optimized for diagnostic dialogue. We set up our consultations in the style of an objective structured clinical examination (OSCE), a practical assessment commonly used in the real world to examine clinicians’ skills and competencies in a standardized and objective way. In a typical OSCE, clinicians might rotate through multiple stations, each simulating a real-life clinical scenario where they perform tasks such as conducting a consultation with a standardized patient actor (trained carefully to emulate a patient with a particular condition). Consultations were performed using a synchronous text-chat tool, mimicking the interface familiar to most consumers using LLMs today.

AMIE is a research AI system based on LLMs for diagnostic reasoning and dialogue.

AMIE: an LLM-based conversational diagnostic research AI system

We trained AMIE on real-world datasets comprising medical reasoning, medical summarization and real-world clinical conversations.

It is feasible to train LLMs using real-world dialogues developed by passively collecting and transcribing in-person clinical visits, however, two substantial challenges limit their effectiveness in training LLMs for medical conversations. First, existing real-world data often fails to capture the vast range of medical conditions and scenarios, hindering the scalability and comprehensiveness. Second, the data derived from real-world dialogue transcripts tends to be noisy, containing ambiguous language (including slang, jargon, humor and sarcasm), interruptions, ungrammatical utterances, and implicit references.

To address these limitations, we designed a self-play based simulated learning environment with automated feedback mechanisms for diagnostic medical dialogue in a virtual care setting, enabling us to scale AMIE’s knowledge and capabilities across many medical conditions and contexts. We used this environment to iteratively fine-tune AMIE with an evolving set of simulated dialogues in addition to the static corpus of real-world data described.

This process consisted of two self-play loops: (1) an “inner” self-play loop, where AMIE leveraged in-context critic feedback to refine its behavior on simulated conversations with an AI patient simulator; and (2) an “outer” self-play loop where the set of refined simulated dialogues were incorporated into subsequent fine-tuning iterations. The resulting new version of AMIE could then participate in the inner loop again, creating a virtuous continuous learning cycle.

Further, we also employed an inference time chain-of-reasoning strategy which enabled AMIE to progressively refine its response conditioned on the current conversation to arrive at an informed and grounded reply.

AMIE uses a novel self-play based simulated dialogue learning environment to improve the quality of diagnostic dialogue across a multitude of disease conditions, specialities and patient contexts.

We tested performance in consultations with simulated patients (played by trained actors), compared to those performed by 20 real PCPs using the randomized approach described above. AMIE and PCPs were assessed from the perspectives of both specialist attending physicians and our simulated patients in a randomized, blinded crossover study that included 149 case scenarios from OSCE providers in Canada, the UK and India in a diverse range of specialties and diseases.

Notably, our study was not designed to emulate either traditional in-person OSCE evaluations or the ways clinicians usually use text, email, chat or telemedicine. Instead, our experiment mirrored the most common way consumers interact with LLMs today, a potentially scalable and familiar mechanism for AI systems to engage in remote diagnostic dialogue.

Overview of the randomized study design to perform a virtual remote OSCE with simulated patients via online multi-turn synchronous text chat.

Performance of AMIE

In this setting, we observed that AMIE performed simulated diagnostic conversations at least as well as PCPs when both were evaluated along multiple clinically-meaningful axes of consultation quality. AMIE had greater diagnostic accuracy and superior performance for 28 of 32 axes from the perspective of specialist physicians, and 24 of 26 axes from the perspective of patient actors.

AMIE outperformed PCPs on multiple evaluation axes for diagnostic dialogue in our evaluations.
Specialist-rated top-k diagnostic accuracy. AMIE and PCPs top-k differential diagnosis (DDx) accuracy are compared across 149 scenarios with respect to the ground truth diagnosis (a) and all diagnoses listed within the accepted differential diagnoses (b). Bootstrapping (n=10,000) confirms all top-k differences between AMIE and PCP DDx accuracy are significant with p <0.05 after false discovery rate (FDR) correction.
Diagnostic conversation and reasoning qualities as assessed by specialist physicians. On 28 out of 32 axes, AMIE outperformed PCPs while being comparable on the rest.

Limitations

Our research has several limitations and should be interpreted with appropriate caution. Firstly, our evaluation technique likely underestimates the real-world value of human conversations, as the clinicians in our study were limited to an unfamiliar text-chat interface, which permits large-scale LLM–patient interactions but is not representative of usual clinical practice. Secondly, any research of this type must be seen as only a first exploratory step on a long journey. Transitioning from a LLM research prototype that we evaluated in this study to a safe and robust tool that could be used by people and those who provide care for them will require significant additional research. There are many important limitations to be addressed, including experimental performance under real-world constraints and dedicated exploration of such important topics as health equity and fairness, privacy, robustness, and many more, to ensure the safety and reliability of the technology.


AMIE as an aid to clinicians

In a recently released preprint, we evaluated the ability of an earlier iteration of the AMIE system to generate a DDx alone or as an aid to clinicians. Twenty (20) generalist clinicians evaluated 303 challenging, real-world medical cases sourced from the New England Journal of Medicine (NEJM) ClinicoPathologic Conferences (CPCs). Each case report was read by two clinicians randomized to one of two assistive conditions: either assistance from search engines and standard medical resources, or AMIE assistance in addition to these tools. All clinicians provided a baseline, unassisted DDx prior to using the respective assistive tools.

Assisted randomized reader study setup to investigate the assistive effect of AMIE to clinicians in solving complex diagnostic case challenges from the New England Journal of Medicine.

AMIE exhibited standalone performance that exceeded that of unassisted clinicians (top-10 accuracy 59.1% vs. 33.6%, p= 0.04). Comparing the two assisted study arms, the top-10 accuracy was higher for clinicians assisted by AMIE, compared to clinicians without AMIE assistance (24.6%, p<0.01) and clinicians with search (5.45%, p=0.02). Further, clinicians assisted by AMIE arrived at more comprehensive differential lists than those without AMIE assistance.

In addition to strong standalone performance, using the AMIE system led to significant assistive effect and improvements in diagnostic accuracy of the clinicians in solving these complex case challenges.

It's worth noting that NEJM CPCs are not representative of everyday clinical practice. They are unusual case reports in only a few hundred individuals so offer limited scope for probing important issues like equity or fairness.


Bold and responsible research in healthcare — the art of the possible

Access to clinical expertise remains scarce around the world. While AI has shown great promise in specific clinical applications, engagement in the dynamic, conversational diagnostic journeys of clinical practice requires many capabilities not yet demonstrated by AI systems. Doctors wield not only knowledge and skill but a dedication to myriad principles, including safety and quality, communication, partnership and teamwork, trust, and professionalism. Realizing these attributes in AI systems is an inspiring challenge that should be approached responsibly and with care. AMIE is our exploration of the “art of the possible”, a research-only system for safely exploring a vision of the future where AI systems might be better aligned with attributes of the skilled clinicians entrusted with our care. It is early experimental-only work, not a product, and has several limitations that we believe merit rigorous and extensive further scientific studies in order to envision a future in which conversational, empathic and diagnostic AI systems might become safe, helpful and accessible.


Acknowledgements

The research described here is joint work across many teams at Google Research and Google Deepmind. We are grateful to all our co-authors - Tao Tu, Mike Schaekermann, Anil Palepu, Daniel McDuff, Jake Sunshine, Khaled Saab, Jan Freyberg, Ryutaro Tanno, Amy Wang, Brenna Li, Mohamed Amin, Sara Mahdavi, Karan Sighal, Shekoofeh Azizi, Nenad Tomasev, Yun Liu, Yong Cheng, Le Hou, Albert Webson, Jake Garrison, Yash Sharma, Anupam Pathak, Sushant Prakash, Philip Mansfield, Shwetak Patel, Bradley Green, Ewa Dominowska, Renee Wong, Juraj Gottweis, Dale Webster, Katherine Chou, Christopher Semturs, Joelle Barral, Greg Corrado and Yossi Matias. We also thank Sami Lachgar, Lauren Winer and John Guilyard for their support with narratives and the visuals. Finally, we are grateful to Michael Howell, James Maynika, Jeff Dean, Karen DeSalvo, Zoubin Gharahmani and Demis Hassabis for their support during the course of this project.

Source: Google AI Blog


Can large language models identify and correct their mistakes?

Posted by Gladys Tyen, Intern, Google Research

LLMs are increasingly popular for reasoning tasks, such as multi-turn QA, task completion, code generation, or mathematics. Yet much like people, they do not always solve problems correctly on the first try, especially on tasks for which they were not trained. Therefore, for such systems to be most useful, they should be able to 1) identify where their reasoning went wrong and 2) backtrack to find another solution.

This has led to a surge in methods related to self-correction, where an LLM is used to identify problems in its own output, and then produce improved results based on the feedback. Self-correction is generally thought of as a single process, but we decided to break it down into two components, mistake finding and output correction.

In “LLMs cannot find reasoning errors, but can correct them!”, we test state-of-the-art LLMs on mistake finding and output correction separately. We present BIG-Bench Mistake, an evaluation benchmark dataset for mistake identification, which we use to address the following questions:

  1. Can LLMs find logical mistakes in Chain-of-Thought (CoT) style reasoning?
  2. Can mistake-finding be used as a proxy for correctness?
  3. Knowing where the mistake is, can LLMs then be prompted to backtrack and arrive at the correct answer?
  4. Can mistake finding as a skill generalize to tasks the LLMs have never seen?

About our dataset

Mistake finding is an underexplored problem in natural language processing, with a particular lack of evaluation tasks in this domain. To best assess the ability of LLMs to find mistakes, evaluation tasks should exhibit mistakes that are non-ambiguous. To our knowledge, most current mistake-finding datasets do not go beyond the realm of mathematics for this reason.

To assess the ability of LLMs to reason about mistakes outside of the math domain, we produce a new dataset for use by the research community, called BIG-Bench Mistake. This dataset consists of Chain-of-Thought traces generated using PaLM 2 on five tasks in BIG-Bench. Each trace is annotated with the location of the first logical mistake.

To maximize the number of mistakes in our dataset, we sample 255 traces where the answer is incorrect (so we know there is definitely a mistake), and 45 traces where the answer is correct (so there may or may not be a mistake). We then ask human labelers to go through each trace and identify the first mistake step. Each trace has been annotated by at least three labelers, whose answers had inter-rater reliability levels of >0.98 (using Krippendorff’s α). The labeling was done for all tasks except the Dyck Languages task, which involves predicting the sequence of closing parentheses for a given input sequence. This task we labeled algorithmically.

The logical errors made in this dataset are simple and unambiguous, providing a good benchmark for testing an LLM’s ability to find its own mistakes before using them on harder, more ambiguous tasks.


Core questions about mistake identification


1. Can LLMs find logical mistakes in Chain-of-Thought style reasoning?

First, we want to find out if LLMs can identify mistakes independently of their ability to correct them. We attempt multiple prompting methods to test GPT series models for their ability to locate mistakes (prompts here) under the assumption that they are generally representative of modern LLM performance.

Generally, we found these state-of-the-art models perform poorly, with the best model achieving 52.9% accuracy overall. Hence, there is a need to improve LLMs’ ability in this area of reasoning.

In our experiments, we try three different prompting methods: direct (trace), direct (step) and CoT (step). In direct (trace), we provide the LLM with the trace and ask for the location step of the mistake or no mistake. In direct (step), we prompt the LLM to ask itself this question for each step it takes. In CoT (step), we prompt the LLM to give its reasoning for whether each step is a mistake or not a mistake.

A diagram showing the three prompting methods direct (trace), direct (step) and CoT (step).

Our finding is in line and builds upon prior results, but goes further in showing that LLMs struggle with even simple and unambiguous mistakes (for comparison, our human raters without prior expertise solve the problem with a high degree of agreement). We hypothesize that this is a big reason why LLMs are unable to self-correct reasoning errors. See the paper for the full results.


2. Can mistake-finding be used as a proxy for correctness of the answer?

When people are confronted with a problem where we are unsure of the answer, we can work through our solutions step-by-step. If no error is found, we can make the assumption that we did the right thing.

While we hypothesized that this would work similarly for LLMs, we discovered that this is a poor strategy. On our dataset of 85% incorrect traces and 15% correct traces, using this method is not much better than the naïve strategy of always labeling traces as incorrect, which gives a weighted average F1 of 78.

A diagram showing how well mistake-finding with LLMs can be used as a proxy for correctness of the answer on each dataset.

3. Can LLMs backtrack knowing where the error is?

Since we’ve shown that LLMs exhibit poor performance in finding reasoning errors in CoT traces, we want to know whether LLMs can even correct errors at all, even if they know where the error is.

Note that knowing the mistake location is different from knowing the right answer: CoT traces can contain logical mistakes even if the final answer is correct, or vice versa. In most real-world situations, we won’t know what the right answer is, but we might be able to identify logical errors in intermediate steps.

We propose the following backtracking method:

  1. Generate CoT traces as usual, at temperature = 0. (Temperature is a parameter that controls the randomness of generated responses, with higher values producing more diverse and creative outputs, usually at the expense of quality.)
  2. Identify the location of the first logical mistake (for example with a classifier, or here we just use labels from our dataset).
  3. Re-generate the mistake step at temperature = 1 and produce a set of eight outputs. Since the original output is known to lead to incorrect results, the goal is to find an alternative generation at this step that is significantly different from the original.
  4. From these eight outputs, select one that is different from the original mistake step. (We just use exact matching here, but in the future this can be something more sophisticated.)
  5. Using the new step, generate the rest of the trace as normal at temperature = 0.

It’s a very simple method that does not require any additional prompt crafting and avoids having to re-generate the entire trace. We test it using the mistake location data from BIG-Bench Mistake, and we find that it can correct CoT errors.

Recent work showed that self-correction methods, like Reflexion and RCI, cause deterioration in accuracy scores because there are more correct answers becoming incorrect than vice versa. Our method, on the other hand, produces more gains (by correcting wrong answers) than losses (by changing right answers to wrong answers).

We also compare our method with a random baseline, where we randomly assume a step to be a mistake. Our results show that this random baseline does produce some gains, but not as much as backtracking with the correct mistake location, and with more losses.

A diagram showing the gains and losses in accuracy for our method as well as a random baseline on each dataset.

4. Can mistake finding generalize to tasks the LLMs have never seen?

To answer this question, we fine-tuned a small model on four of the BIG-Bench tasks and tested it on the fifth, held-out task. We do this for every task, producing five fine-tuned models in total. Then we compare the results with just zero-shot prompting PaLM 2-L-Unicorn, a much larger model.

Bar chart showing the accuracy improvement of the fine-tuned small model compared to zero-shot prompting with PaLM 2-L-Unicorn.

Our results show that the much smaller fine-tuned reward model generally performs better than zero-shot prompting a large model, even though the reward model has never seen data from the task in the test set. The only exception is logical deduction, where it performs on par with zero-shot prompting.

This is a very promising result as we can potentially just use a small fine-tuned reward model to perform backtracking and improve accuracy on any task, even if we don’t have the data for it. This smaller reward model is completely independent of the generator LLM, and can be updated and further fine-tuned for individual use cases.

An illustration showing how our backtracking method works.

Conclusion

In this work, we created an evaluation benchmark dataset that the wider academic community can use to evaluate future LLMs. We further showed that LLMs currently struggle to find logical errors. However, if they could, we show the effectiveness of backtracking as a strategy that can provide gains on tasks. Finally, a smaller reward model can be trained on general mistake-finding tasks and be used to improve out-of-domain mistake finding, showing that mistake-finding can generalize.


Acknowledgements

Thank you to Peter Chen, Tony Mak, Hassan Mansoor and Victor Cărbune for contributing ideas and helping with the experiments and data collection. We would also like to thank Sian Gooding and Vicky Zayats for their comments and suggestions on the paper.


Source: Google AI Blog


Responsible AI at Google Research: User Experience Team

Posted by Ayça Çakmakli, UX Lead, Google Research, Responsible AI and Human Centered Technology Team

Google’s Responsible AI User Experience (Responsible AI UX) team is a product-minded team embedded within Google Research. This unique positioning requires us to apply responsible AI development practices to our user-centered user experience (UX) design process. In this post, we describe the importance of UX design and responsible AI in product development, and share a few examples of how our team’s capabilities and cross-functional collaborations have led to responsible development across Google.

First, the UX part. We are a multi-disciplinary team of product design experts: designers, engineers, researchers, and strategists who manage the user-centered UX design process from early-phase ideation and problem framing to later-phase user-interface (UI) design, prototyping and refinement. We believe that effective product development occurs when there is clear alignment between significant unmet user needs and a product's primary value proposition, and that this alignment is reliably achieved via a thorough user-centered UX design process.

And second, recognizing generative AI’s (GenAI) potential to significantly impact society, we embrace our role as the primary user advocate as we continue to evolve our UX design process to meet the unique challenges AI poses, maximizing the benefits and minimizing the risks. As we navigate through each stage of an AI-powered product design process, we place a heightened emphasis on the ethical, societal, and long-term impact of our decisions. We contribute to the ongoing development of comprehensive safety and inclusivity protocols that define design and deployment guardrails around key issues like content curation, security, privacy, model capabilities, model access, equitability, and fairness that help mitigate GenAI risks.

Responsible AI UX is constantly evolving its user-centered product design process to meet the needs of a GenAI-powered product landscape with greater sensitivity to the needs of users and society and an emphasis on ethical, societal, and long-term impact.

Responsibility in product design is also reflected in the user and societal problems we choose to address and the programs we resource. Thus, we encourage the prioritization of user problems with significant scale and severity to help maximize the positive impact of GenAI technology.

Communication across teams and disciplines is essential to responsible product design. The seamless flow of information and insight from user research teams to product design and engineering teams, and vice versa, is essential to good product development. One of our team’s core objectives is to ensure the practical application of deep user-insight into AI-powered product design decisions at Google by bridging the communication gap between the vast technological expertise of our engineers and the user/societal expertise of our academics, research scientists, and user-centered design research experts. We’ve built a multidisciplinary team with expertise in these areas, deepening our empathy for the communication needs of our audience, and enabling us to better interface between our user & society experts and our technical experts. We create frameworks, guidebooks, prototypes, cheatsheets, and multimedia tools to help bring insights to life for the right people at the right time.



Facilitating responsible GenAI prototyping and development

During collaborations between Responsible AI UX, the People + AI Research (PAIR) initiative and Labs, we identified that prototyping can afford a creative opportunity to engage with large language models (LLM), and is often the first step in GenAI product development. To address the need to introduce LLMs into the prototyping process, we explored a range of different prompting designs. Then, we went out into the field, employing various external, first-person UX design research methodologies to draw out insight and gain empathy for the user’s perspective. Through user/designer co-creation sessions, iteration, and prototyping, we were able to bring internal stakeholders, product managers, engineers, writers, sales, and marketing teams along to ensure that the user point of view was well understood and to reinforce alignment across teams.

The result of this work was MakerSuite, a generative AI platform launched at Google I/O 2023 that enables people, even those without any ML experience, to prototype creatively using LLMs. The team’s first-hand experience with users and understanding of the challenges they face allowed us to incorporate our AI Principles into the MakerSuite product design. Product features like safety filters, for example, enable users to manage outcomes, leading to easier and more responsible product development with MakerSuite.

Because of our close collaboration with product teams, we were able to adapt text-only prototyping to support multimodal interaction with Google AI Studio, an evolution of MakerSuite. Now, Google AI Studio enables developers and non-developers alike to seamlessly leverage Google’s latest Gemini model to merge multiple modality inputs, like text and image, in product explorations. Facilitating product development in this way provides us with the opportunity to better use AI to identify appropriateness of outcomes and unlocks opportunities for developers and non-developers to play with AI sandboxes. Together with our partners, we continue to actively push this effort in the products we support.

Google AI studio enables developers and non-developers to leverage Google Cloud infrastructure and merge multiple modality inputs in their product explorations.


Equitable speech recognition

Multiple external studies, as well as Google’s own research, have identified an unfortunate deficiency in the ability of current speech recognition technology to understand Black speakers on average, relative to White speakers. As multimodal AI tools begin to rely more heavily on speech prompts, this problem will grow and continue to alienate users. To address this problem, the Responsible AI UX team is partnering with world-renowned linguists and scientists at Howard University, a prominent HBCU, to build a high quality African-American English dataset to improve the design of our speech technology products to make them more accessible. Called Project Elevate Black Voices, this effort will allow Howard University to share the dataset with those looking to improve speech technology while establishing a framework for responsible data collection, ensuring the data benefits Black communities. Howard University will retain the ownership and licensing of the dataset and serve as stewards for its responsible use. At Google, we’re providing funding support and collaborating closely with our partners at Howard University to ensure the success of this program.




Equitable computer vision

The Gender Shades project highlighted that computer vision systems struggle to detect people with darker skin tones, and performed particularly poorly for women with darker skin tones. This is largely due to the fact that the datasets used to train these models were not inclusive to a wide range of skin tones. To address this limitation, the Responsible AI UX team has been partnering with sociologist Dr. Ellis Monk to release the Monk Skin Tone Scale (MST), a skin tone scale designed to be more inclusive of the spectrum of skin tones around the world. It provides a tool to assess the inclusivity of datasets and model performance across an inclusive range of skin tones, resulting in features and products that work better for everyone.

We have integrated MST into a range of Google products, such as Search, Google Photos, and others. We also open sourced MST, published our research, described our annotation practices, and shared an example dataset to encourage others to easily integrate it into their products. The Responsible AI UX team continues to collaborate with Dr. Monk, utilizing the MST across multiple product applications and continuing to do international research to ensure that it is globally inclusive.


Consulting & guidance

As teams across Google continue to develop products that leverage the capabilities of GenAI models, our team recognizes that the challenges they face are varied and that market competition is significant. To support teams, we develop actionable assets to facilitate a more streamlined and responsible product design process that considers available resources. We act as a product-focused design consultancy, identifying ways to scale services, share expertise, and apply our design principles more broadley. Our goal is to help all product teams at Google connect significant unmet user needs with technology benefits via great responsible product design.

One way we have been doing this is with the creation of the People + AI Guidebook, an evolving summative resource of many of the responsible design lessons we’ve learned and recommendations we’ve made for internal and external stakeholders. With its forthcoming, rolling updates focusing specifically on how to best design and consider user needs with GenAI, we hope that our internal teams, external stakeholders, and larger community will have useful and actionable guidance at the most critical milestones in the product development journey.

The People + AI Guidebook has six chapters, designed to cover different aspects of the product life cycle.

If you are interested in reading more about Responsible AI UX and how we are specifically thinking about designing responsibly with Generative AI, please check out this Q&A piece.


Acknowledgements

Shout out to our the Responsible AI UX team members: Aaron Donsbach, Alejandra Molina, Courtney Heldreth, Diana Akrong, Ellis Monk, Femi Olanubi, Hope Neveux, Kafayat Abdul, Key Lee, Mahima Pushkarna, Sally Limb, Sarah Post, Sures Kumar Thoddu Srinivasan, Tesh Goyal, Ursula Lauriston, and Zion Mengesha. Special thanks to Michelle Cohn for her contributions to this work.

Source: Google AI Blog