Tag Archives: Computer Vision

Foundation models for reasoning on charts

Visual language is the form of communication that relies on pictorial symbols outside of text to convey information. It is ubiquitous in our digital life in the form of iconography, infographics, tables, plots, and charts, extending to the real world in street signs, comic books, food labels, etc. For that reason, having computers better understand this type of media can help with scientific communication and discovery, accessibility, and data transparency.

While computer vision models have made tremendous progress using learning-based solutions since the advent of ImageNet, the focus has been on natural images, where all sorts of tasks, such as classification, visual question answering (VQA), captioning, detection and segmentation, have been defined, studied and in some cases advanced to reach human performance. However, visual language has not garnered a similar level of attention, possibly because of the lack of large-scale training sets in this space. But over the last few years, new academic datasets have been created with the goal of evaluating question answering systems on visual language images, like PlotQA, InfographicsVQA, and ChartQA.

Example from ChartQA. Answering the question requires reading the information and computing the sum and the difference.

Existing models built for these tasks relied on integrating optical character recognition (OCR) information and their coordinates into larger pipelines but the process is error prone, slow, and generalizes poorly. The prevalence of these methods was because existing end-to-end computer vision models based on convolutional neural networks (CNNs) or transformers pre-trained on natural images could not be easily adapted to visual language. But existing models are ill-prepared for the challenges in answering questions on charts, including reading the relative height of bars or the angle of slices in pie charts, understanding axis scales, correctly mapping pictograms with their legend values with colors, sizes and textures, and finally performing numerical operations with the extracted numbers.

In light of these challenges, we propose “MatCha: Enhancing Visual Language Pretraining with Math Reasoning and Chart Derendering”. MatCha, which stands for math and charts, is a pixels-to-text foundation model (a pre-trained model with built-in inductive biases that can be fine-tuned for multiple applications) trained on two complementary tasks: (a) chart de-rendering and (b) math reasoning. In chart de-rendering, given a plot or chart, the image-to-text model is required to generate its underlying data table or the code used to render it. For math reasoning pre-training, we pick textual numerical reasoning datasets and render the input into images, which the image-to-text model needs to decode for answers. We also propose “DePlot: One-shot visual language reasoning by plot-to-table translation”, a model built on top of MatCha for one-shot reasoning on charts via translation to tables. With these methods we surpass the previous state of the art in ChartQA by more than 20% and match the best summarization systems that have 1000 times more parameters. Both papers will be presented at ACL2023.

Chart de-rendering

Plots and charts are usually generated by an underlying data table and a piece of code. The code defines the overall layout of the figure (e.g., type, direction, color/shape scheme) and the underlying data table establishes the actual numbers and their groupings. Both the data and code are sent to a compiler/rendering engine to create the final image. To understand a chart, one needs to discover the visual patterns in the image and effectively parse and group them to extract the key information. Reversing the plot rendering process demands all such capabilities and can thus serve as an ideal pre-training task.

A chart created from a table in the Airbus A380 Wikipedia page using random plotting options. The pre-training task for MatCha consists of recovering the source table or the source code from the image.

In practice, it is challenging to simultaneously obtain charts, their underlying data tables, and their rendering code. To collect sufficient pre-training data, we independently accumulate [chart, code] and [chart, table] pairs. For [chart, code], we crawl all GitHub IPython notebooks with appropriate licenses and extract blocks with figures. A figure and the code block right before it are saved as a [chart, code] pair. For [chart, table] pairs, we explored two sources. For the first source, synthetic data, we manually write code to convert web-crawled Wikipedia tables from the TaPas codebase to charts. We sampled from and combined several plotting options depending on the column types. In addition, we also add [chart, table] pairs generated in PlotQA to diversify the pre-training corpus. The second source is web-crawled [chart, table] pairs. We directly use the [chart, table] pairs crawled in the ChartQA training set, containing around 20k pairs in total from four websites: Statista, Pew, Our World in Data, and OECD.

Math reasoning

We incorporate numerical reasoning knowledge into MatCha by learning math reasoning skills from textual math datasets. We use two existing textual math reasoning datasets, MATH and DROP for pre-training. MATH is synthetically created, containing two million training examples per module (type) of questions. DROP is a reading-comprehension–style QA dataset where the input is a paragraph context and a question.

To solve questions in DROP, the model needs to read the paragraph, extract relevant numbers and perform numerical computation. We found both datasets to be complementary. MATH contains a large number of questions across different categories, which helps us identify math operations needed to explicitly inject into the model. DROP’s reading-comprehension format resembles the typical QA format wherein models simultaneously perform information extraction and reasoning. In practice, we render inputs of both datasets into images. The model is trained to decode the answer.

To improve the math reasoning skills of MatCha we incorporate examples from MATH and DROP into the pre-training objective, by rendering the input text as images.

End-to-end results

We use a Pix2Struct model backbone, which is an image-to-text transformer tailored for website understanding, and pre-train it with the two tasks described above. We demonstrate the strengths of MatCha by fine-tuning it on several visual language tasks — tasks involving charts and plots for question answering and summarization where no access to the underlying table is possible. MatCha surpasses previous models’ performance by a large margin and also outperforms the previous state of the art, which assumes access to underlying tables.

In the figure below, we first evaluate two baseline models that incorporate information from an OCR pipeline, which until recently was the standard approach for working with charts. The first is based on T5, the second on VisionTaPas. We also compare against PaLI-17B, which is a large (~1000 times larger than the other models) image plus text-to-text transformer trained on a diverse set of tasks but with limited capabilities for reading text and other forms of visual language. Finally, we report the Pix2Struct and MatCha model results.

Experimental results on two chart QA benchmarks ChartQA & PlotQA (using relaxed accuracy) and a chart summarization benchmark chart-to-text (using BLEU4). Matcha surpasses the state of the art by a large margin on QA, compared to larger models, and matches these larger models on summarization.

For QA datasets, we use the official relaxed accuracy metric that allows for small relative errors in numerical outputs. For chart-to-text summarization, we report BLEU scores. MatCha achieves noticeably improved results compared to baselines for question answering, and comparable results to PaLI in summarization, where large size and extensive long text/captioning generation pre-training are advantageous for this kind of long-form text generation.

Derendering plus large language model chains

While extremely performant for their number of parameters, particularly on extractive tasks, we observed that fine-tuned MatCha models could still struggle with end-to-end complex reasoning (e.g., mathematical operations involving large numbers or multiple steps). Thus, we also propose a two-step method to tackle this: 1) a model reads a chart, then outputs the underlying table, 2) a large language model (LLM) reads this output and then tries to answer the question solely based on the textual input.

For the first model, we fine-tuned MatCha solely on the chart-to-table task, increasing the output sequence length to guarantee it could recover all or most of the information in the chart. DePlot is the resulting model. In the second stage, any LLM (such as FlanPaLM or Codex) can be used for the task, and we can rely on the standard methods to increase performance on LLMs, for example chain-of-thought and self-consistency. We also experimented with program-of-thoughts where the model produces executable Python code to offload complex computations.

An illustration of the DePlot+LLM method. This is a real example using FlanPaLM and Codex. The blue boxes are input to the LLM and the red boxes contain the answer generated by the LLMs. We highlight some of the key reasoning steps in each answer.

As shown in the example above, the DePlot model in combination with LLMs outperforms fine-tuned models by a significant margin, especially so in the human-sourced portion of ChartQA, where the questions are more natural but demand more difficult reasoning. Furthermore, DePlot+LLM can do so without access to any training data.

We have released the new models and code at our GitHub repo, where you can try it out yourself in colab. Checkout the papers for MatCha and DePlot for more details on the experimental results. We hope that our results can benefit the research community and make the information in charts and plots more accessible to everyone.


This work was carried out by Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen and Yasemin Altun from our Language Team as part of Fangyu's internship project. Nigel Collier from Cambridge also was a collaborator. We would like to thank Joshua Howland, Alex Polozov, Shrestha Basu Mallick, Massimo Nicosia and William Cohen for their valuable comments and suggestions.

Source: Google AI Blog

Sparse video tubes for joint video and image vision transformers

Video understanding is a challenging problem that requires reasoning about both spatial information (e.g., for objects in a scene, including their locations and relations) and temporal information for activities or events shown in a video. There are many video understanding applications and tasks, such as understanding the semantic content of web videos and robot perception. However, current works, such as ViViT and TimeSFormer, densely process the video and require significant compute, especially as model size plus video length and resolution increase.

In “Rethinking Video ViTs: Sparse Video Tubes for Joint Image and Video Learning”, to be presented at CVPR 2023, we introduce a simple technique that turns a Vision Transformer (ViT) model image encoder into an efficient video backbone using sparse video tubes (learnable visual representations of samples from the video) to reduce the model’s compute needs. This approach can seamlessly process both images and videos, which allows it to leverage both image and video data sources during training. This training further enables our sparse tubes ViT model to coalesce image and video backbones together to serve a dual role as either an image or video backbone (or both), depending on the input. We demonstrate that this model is scalable, can be adapted to large pre-trained ViTs without requiring full fine-tuning, and achieves state-of-the-art results across many video classification benchmarks.

Using sparse video tubes to sample a video, combined with a standard ViT encoder, leads to an efficient visual representation that can be seamlessly shared with image inputs.

Building a joint image-video backbone

Our sparse tube ViT uses a standard ViT backbone, consisting of a stack of Transformer layers, that processes video information. Previous methods, such as ViViT, densely tokenize the video and then apply factorized attention, i.e., the attention weights for each token are computed separately for the temporal and spatial dimensions. In the standard ViT architecture, self-attention is computed over the whole token sequence. When using videos as input, token sequences become quite long, which can make this computation slow. Instead, in the method we propose, the video is sparsely sampled using video tubes, which are 3D learnable visual representations of various shapes and sizes (described in more detail below) from the video. These tubes are used to sparsely sample the video using a large temporal stride, i.e., when a tube kernel is only applied to a few locations in the video, rather than every pixel.

By sparsely sampling the video tubes, we can use the same global self-attention module, rather than factorized attention like ViViT. We experimentally show that the addition of factorized attention layers can harm the performance due to the uninitialized weights. This single stack of transformer layers in the ViT backbone also enables better sharing of the weights and improves performance. Sparse video tube sampling is done by using a large spatial and temporal stride that selects tokens on a fixed grid. The large stride reduces the number of tokens in the full network, while still capturing both spatial and temporal information and enabling the efficient processing of all tokens.

Sparse video tubes

Video tubes are 3D grid-based cuboids that can have different shapes or categories and capture different information with strides and starting locations that can overlap. In the model, we use three distinct tube shapes that capture: (1) only spatial information (resulting in a set of 2D image patches), (2) long temporal information (over a small spatial area), and (3) both spatial and temporal information equally. Tubes that capture only spatial information can be applied to both image and video inputs. Tubes that capture long temporal information or both temporal and spatial information equally are only applied to video inputs. Depending on the input video size, the three tube shapes are applied to the model multiple times to generate tokens.

A fixed position embedding, which captures the global location of each tube (including any strides, offsets, etc.) relative to all the other tubes, is applied to the video tubes. Different from the previous learned position embeddings, this fixed one better enables sparse, overlapping sampling. Capturing the global location of the tube helps the model know where each came from, which is especially helpful when tubes overlap or are sampled from distant video locations. Next, the tube features are concatenated together to form a set of N tokens. These tokens are processed by a standard ViT encoder. Finally, we apply an attention pooling to compress all the tokens into a single representation and input to a fully connected (FC) layer to make the classification (e.g., playing soccer, swimming, etc.).

Our video ViT model works by sampling sparse video tubes from the video (shown at the bottom) to enable either or both image or video inputs to be seamlessly processed. These tubes have different shapes and capture different video features. Tube 1 (yellow) only captures spatial information, resulting in a set of 2D patches that can be applied to image inputs. Tube 2 (red) captures temporal information and some spatial information and tube 3 (green) equally captures both temporal and spatial information (i.e., the spatial size of the tube x and y are the same as the number of frames t). Tubes 2 and 3 can only be applied to video inputs. The position embedding is added to all the tube features.

Scaling video ViTs

The process of building video backbones is computationally intensive, but our sparse tube ViT model enables computationally efficient scaling of video models, leveraging previously trained image backbones. Since image backbones can be adapted to a video backbone, large image backbones can be turned into large video backbones. More specifically, one can transfer the learned video feature representations from a small tube ViT to a large pre-trained image ViT and train the resulting model with video data for only a few steps, as opposed to a full training from scratch.

Our approach enables scaling a sparse tube ViT in a more efficient way. Specifically, the video features from a small video ViT (top network) can be transferred to a large, pre-trained image ViT (bottom network), and further fine-tuned. This requires fewer training steps to achieve strong performance with the large model. This is beneficial as large video models might be prohibitively expensive to train from scratch.


We evaluate our sparse tube ViT approach using Kinetics-400 (shown below), Kinetics-600 and Kinetics-700 datasets and compare its performance to a long list of prior methods. We find that our approach outperforms all prior methods. Importantly, it outperforms all state-of-the-art methods trained jointly on image+video datasets.

Performance compared to several prior works on the popular Kinetics-400 video dataset. Our sparse tube ViT outperforms state-of-the-art methods.

Furthermore, we test our sparse tube ViT model on the Something-Something V2 dataset, which is commonly used to evaluate more dynamic activities, and also report that it outperforms all prior state-of-the-art approaches.

Performance on the Something-Something V2 video dataset.

Visualizing some learned kernels

It is interesting to understand what kind of rudimentary features are being learned by the proposed model. We visualize them below, showing both the 2D patches, which are shared for both images and videos, and video tubes. These visualizations show the 2D or 3D information being captured by the projection layer. For example, in the 2D patches, various common features, like edges and colors, are detected, while the 3D tubes capture basic shapes and how they may change over time.

Visualizations of patches and tubes learned the sparse tube ViT model. Top row are the 2D patches and the remaining two rows are snapshots from the learned video tubes. The tubes show each patch for the 8 or 4 frames to which they are applied.


We have presented a new sparse tube ViT, which can turn a ViT encoder into an efficient video model, and can seamlessly work with both image and video inputs. We also showed that large video encoders can be bootstrapped from small video encoders and image-only ViTs. Our approach outperforms prior methods across several popular video understanding benchmarks. We believe that this simple representation can facilitate much more efficient learning with input videos, seamlessly incorporate either image or video inputs and effectively eliminate the bifurcation of image and video models for future multimodal understanding.


This work is conducted by AJ Piergiovanni, Weicheng Kuo and Anelia Angelova, who are now at Google DeepMind. We thank Abhijit Ogale, Luowei Zhou, Claire Cui and our colleagues in Google Research for their helpful discussions, comments, and support.

Source: Google AI Blog

F-VLM: Open-vocabulary object detection upon frozen vision and language models

Detection is a fundamental vision task that aims to localize and recognize objects in an image. However, the data collection process of manually annotating bounding boxes or instance masks is tedious and costly, which limits the modern detection vocabulary size to roughly 1,000 object classes. This is orders of magnitude smaller than the vocabulary people use to describe the visual world and leaves out many categories. Recent vision and language models (VLMs), such as CLIP, have demonstrated improved open-vocabulary visual recognition capabilities through learning from Internet-scale image-text pairs. These VLMs are applied to zero-shot classification using frozen model weights without the need for fine-tuning, which stands in stark contrast to the existing paradigms used for retraining or fine-tuning VLMs for open-vocabulary detection tasks.

Intuitively, to align the image content with the text description during training, VLMs may learn region-sensitive and discriminative features that are transferable to object detection. Surprisingly, features of a frozen VLM contain rich information that are both region sensitive for describing object shapes (second column below) and discriminative for region classification (third column below). In fact, feature grouping can nicely delineate object boundaries without any supervision. This motivates us to explore the use of frozen VLMs for open-vocabulary object detection with the goal to expand detection beyond the limited set of annotated categories.

We explore the potential of frozen vision and language features for open-vocabulary detection. The K-Means feature grouping reveals rich semantic and region-sensitive information where object boundaries are nicely delineated (column 2). The same frozen features can classify groundtruth (GT) regions well without fine-tuning (column 3).

In “F-VLM: Open-Vocabulary Object Detection upon Frozen Vision and Language Models”, presented at ICLR 2023, we introduce a simple and scalable open-vocabulary detection approach built upon frozen VLMs. F-VLM reduces the training complexity of an open-vocabulary detector to below that of a standard detector, obviating the need for knowledge distillation, detection-tailored pre-training, or weakly supervised learning. We demonstrate that by preserving the knowledge of pre-trained VLMs completely, F-VLM maintains a similar philosophy to ViTDet and decouples detector-specific learning from the more task-agnostic vision knowledge in the detector backbone. We are also releasing the F-VLM code along with a demo on our project page.

Learning upon frozen vision and language models

We desire to retain the knowledge of pretrained VLMs as much as possible with a view to minimize effort and cost needed to adapt them for open-vocabulary detection. We use a frozen VLM image encoder as the detector backbone and a text encoder for caching the detection text embeddings of offline dataset vocabulary. We take this VLM backbone and attach a detector head, which predicts object regions for localization and outputs detection scores that indicate the probability of a detected box being of a certain category. The detection scores are the cosine similarity of region features (a set of bounding boxes that the detector head outputs) and category text embeddings. The category text embeddings are obtained by feeding the category names through the text model of pretrained VLM (which has both image and text models)r.

The VLM image encoder consists of two parts: 1) a feature extractor and 2) a feature pooling layer. We adopt the feature extractor for detector head training, which is the only step we train (on standard detection data), to allow us to directly use frozen weights, inheriting rich semantic knowledge (e.g., long-tailed categories like martini, fedora hat, pennant) from the VLM backbone. The detection losses include box regression and classification losses.

At training time, F-VLM is simply a detector with the last classification layer replaced by base-category text embeddings.

Region-level open-vocabulary recognition

The ability to perform open-vocabulary recognition at region level (i.e., bounding box level as opposed to image level) is integral to F-VLM. Since the backbone features are frozen, they do not overfit to the training categories (e.g., donut, zebra) and can be directly cropped for region-level classification. F-VLM performs this open-vocabulary classification only at test time. To obtain the VLM features for a region, we apply the feature pooling layer on the cropped backbone output features. Because the pooling layer requires fixed-size inputs, e.g., 7x7 for ResNet50 (R50) CLIP backbone, we crop and resize the region features with the ROI-Align layer (shown below). Unlike existing open-vocabulary detection approaches, we do not crop and resize the RGB image regions and cache their embeddings in a separate offline process, but train the detector head in one stage. This is simpler and makes more efficient use of disk storage space.. In addition, we do not crop VLM region features during training because the backbone features are frozen.

Despite never being trained on regions, the cropped region features maintain good open-vocabulary recognition capability. However, we observe the cropped region features are not sensitive enough to the localization quality of the regions, i.e., a loosely vs. tightly localized box both have similar features. This may be good for classification, but is problematic for detection because we need the detection scores to reflect localization quality as well. To remedy this, we apply the geometric mean to combine the VLM scores with the detection scores for each region and category. The VLM scores indicate the probability of a detection box being of a certain category according to the pretrained VLM. The detection scores indicate the class probability distribution of each box based on the similarity of region features and input text embeddings.

At test time, F-VLM uses the region proposals to crop out the top-level features of the VLM backbone and compute the VLM score per region. The trained detector head provides the detection boxes and masks, while the final detection scores are a combination of detection and VLM scores.


We apply F-VLM to the popular LVIS open-vocabulary detection benchmark. At the system-level, the best F-VLM achieves 32.8 average precision (AP) on rare categories (APr), which outperforms the state of the art by 6.5 mask APr and many other approaches based on knowledge distillation, pre-training, or joint training with weak supervision. F-VLM shows strong scaling property with frozen model capacity, while the number of trainable parameters is fixed. Moreover, F-VLM generalizes and scales well in the transfer detection tasks (e.g., Objects365 and Ego4D datasets) by simply replacing the vocabularies without fine-tuning the model. We test the LVIS-trained models on the popular Objects365 datasets and demonstrate that the model can work very well without training on in-domain detection data.

F-VLM outperforms the state of the art (SOTA) on LVIS open-vocabulary detection benchmark and transfer object detection. On the x-axis, we show the LVIS metric mask AP on rare categories (APr), and the Objects365 (O365) metric box AP on all categories. The sizes of the detector backbones are as follows: Small(R50), Base (R50x4), Large(R50x16), Huge(R50x64). The naming follows CLIP convention.

We visualize F-VLM on open-vocabulary detection and transfer detection tasks (shown below). On LVIS and Objects365, F-VLM correctly detects both novel and common objects. A key benefit of open-vocabulary detection is to test on out-of-distribution data with categories given by users on the fly. See the F-VLM paper for more visualization on LVIS, Objects365 and Ego4D datasets.

F-VLM open-vocabulary and transfer detections. Top: Open-vocabulary detection on LVIS. We only show the novel categories for clarity. Bottom: Transfer to Objects365 dataset shows accurate detection of many categories. Novel categories detected: fedora, martini, pennant, football helmet (LVIS); slide (Objects365).

Training efficiency

We show that F-VLM can achieve top performance with much less computational resources in the table below. Compared to the state-of-the-art approach, F-VLM can achieve better performance with 226x fewer resources and 57x faster wall clock time. Apart from training resource savings, F-VLM has potential for substantial memory savings at training time by running the backbone in inference mode. The F-VLM system runs almost as fast as a standard detector at inference time, because the only addition is a single attention pooling layer on the detected region features.

Method       APr       Training Epochs       Training Cost
      Training Cost Savings      
SOTA       26.3       460       8,000       1x      
F-VLM       32.8       118       565       14x      
F-VLM       31.0       14.7       71       113x      
F-VLM       27.7       7.4       35       226x      

We provide additional results using the shorter Detectron2 training recipes (12 and 36 epochs), and show similarly strong performance by using a frozen backbone. The default setting is marked in gray.

Backbone       Large Scale Jitter       #Epochs       Batch Size       APr      
R50             12       16       18.1      
R50             36       64       18.5      
R50             100       256       18.6      
R50x64             12       16       31.9      
R50x64             36       64       32.6      
R50x64             100       256       32.8      


We present F-VLM – a simple open-vocabulary detection method which harnesses the power of frozen pre-trained large vision-language models to provide detection of novel objects. This is done without a need for knowledge distillation, detection-tailored pre-training, or weakly supervised learning. Our approach offers significant compute savings and obviates the need for image-level labels. F-VLM achieves the new state-of-the-art in open-vocabulary detection on the LVIS benchmark at system level, and shows very competitive transfer detection on other datasets. We hope this study can both facilitate further research in novel-object detection and help the community explore frozen VLMs for a wider range of vision tasks.


This work is conducted by Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. We would like to thank our colleagues at Google Research for their advice and helpful discussions.

Source: Google AI Blog

MaMMUT: A simple vision-encoder text-decoder architecture for multimodal tasks

Vision-language foundational models are built on the premise of a single pre-training followed by subsequent adaptation to multiple downstream tasks. Two main and disjoint training scenarios are popular: a CLIP-style contrastive learning and next-token prediction. Contrastive learning trains the model to predict if image-text pairs correctly match, effectively building visual and text representations for the corresponding image and text inputs, whereas next-token prediction predicts the most likely next text token in a sequence, thus learning to generate text, according to the required task. Contrastive learning enables image-text and text-image retrieval tasks, such as finding the image that best matches a certain description, and next-token learning enables text-generative tasks, such as Image Captioning and Visual Question Answering (VQA). While both approaches have demonstrated powerful results, when a model is pre-trained contrastively, it typically does not fare well on text-generative tasks and vice-versa. Furthermore, adaptation to other tasks is often done with complex or inefficient methods. For example, in order to extend a vision-language model to videos, some models need to do inference for each video frame separately. This limits the size of the videos that can be processed to only a few frames and does not fully take advantage of motion information available across frames.

Motivated by this, we present “A Simple Architecture for Joint Learning for MultiModal Tasks”, called MaMMUT, which is able to train jointly for these competing objectives and which provides a foundation for many vision-language tasks either directly or via simple adaptation. MaMMUT is a compact, 2B-parameter multimodal model that trains across contrastive, text generative, and localization-aware objectives. It consists of a single image encoder and a text decoder, which allows for a direct reuse of both components. Furthermore, a straightforward adaptation to video-text tasks requires only using the image encoder once and can handle many more frames than prior work. In line with recent language models (e.g., PaLM, GLaM, GPT3), our architecture uses a decoder-only text model and can be thought of as a simple extension of language models. While modest in size, our model outperforms the state of the art or achieves competitive performance on image-text and text-image retrieval, video question answering (VideoQA), video captioning, open-vocabulary detection, and VQA.

The MaMMUT model enables a wide range of tasks such as image-text/text-image retrieval (top left and top right), VQA (middle left), open-vocabulary detection (middle right), and VideoQA (bottom).

Decoder-only model architecture

One surprising finding is that a single language-decoder is sufficient for all these tasks, which obviates the need for both complex constructs and training procedures presented before. For example, our model (presented to the left in the figure below) consists of a single visual encoder and single text-decoder, connected via cross attention, and trains simultaneously on both contrastive and text-generative types of losses. Comparatively, prior work is either not able to handle image-text retrieval tasks, or applies only some losses to only some parts of the model. To enable multimodal tasks and fully take advantage of the decoder-only model, we need to jointly train both contrastive losses and text-generative captioning-like losses.

MaMMUT architecture (left) is a simple construct consisting of a single vision encoder and a single text decoder. Compared to other popular vision-language models — e.g., PaLI (middle) and ALBEF, CoCa (right) — it trains jointly and efficiently for multiple vision-language tasks, with both contrastive and text-generative losses, fully sharing the weights between the tasks.

Decoder two-pass learning

Decoder-only models for language learning show clear advantages in performance with smaller model size (almost half the parameters). The main challenge for applying them to multimodal settings is to unify the contrastive learning (which uses unconditional sequence-level representation) with captioning (which optimizes the likelihood of a token conditioned on the previous tokens). We propose a two-pass approach to jointly learn these two conflicting types of text representations within the decoder. During the first pass, we utilize cross attention and causal masking to learn the caption generation task — the text features can attend to the image features and predict the tokens in sequence. On the second pass, we disable the cross-attention and causal masking to learn the contrastive task. The text features will not see the image features but can attend bidirectionally to all text tokens at once to produce the final text-based representation. Completing this two-pass approach within the same decoder allows for accommodating both types of tasks that were previously hard to reconcile. While simple, we show that this model architecture is able to provide a foundation for multiple multimodal tasks.

MaMMUT decoder-only two-pass learning enables both contrastive and generative learning paths by the same model.

Another advantage of our architecture is that, since it is trained for these disjoint tasks, it can be seamlessly applied to multiple applications such as image-text and text-image retrieval, VQA, and captioning.

Moreover, MaMMUT easily adapts to video-language tasks. Previous approaches used a vision encoder to process each frame individually, which required applying it multiple times. This is slow and restricts the number of frames the model can handle, typically to only 6–8. With MaMMUT, we use sparse video tubes for lightweight adaptation directly via the spatio-temporal information from the video. Furthermore, adapting the model to Open-Vocabulary Detection is done by simply training to detect bounding-boxes via an object-detection head.

Adaptation of the MaMMUT architecture to video tasks (left) is simple and fully reuses the model. This is done by generating a video “tubes” feature representation, similar to image patches, that are projected to lower dimensional tokens and run through the vision encoder. Unlike prior approaches (right) that need to run multiple individual images through the vision encoder, we use it only once.


Our model achieves excellent zero-shot results on image-text and text-image retrieval without any adaptation, outperforming all previous state-of-the-art models. The results on VQA are competitive with state-of-the-art results, which are achieved by much larger models. The PaLI model (17B parameters) and the Flamingo model (80B) have the best performance on the VQA2.0 dataset, but MaMMUT (2B) has the same accuracy as the 15B PaLI.

MaMMUT outperforms the state of the art (SOTA) on Zero-Shot Image-Text (I2T) and Text-Image (T2I) retrieval on both MS-COCO (top) and Flickr (bottom) benchmarks.
Performance on the VQA2.0 dataset is competitive but does not outperform large models such as Flamingo-80B and PalI-17B. Performance is evaluated in the more challenging open-ended text generation setting.

MaMMUT also outperforms the state-of-the-art on VideoQA, as shown below on the MSRVTT-QA and MSVD-QA datasets. Note that we outperform much bigger models such as Flamingo, which is specifically designed for image+video pre-training and is pre-trained with both image-text and video-text data.

MaMMUT outperforms the SOTA models on VideoQA tasks (MSRVTT-QA dataset, top, MSVD-QA dataset, bottom), outperforming much larger models, e.g., the 5B GIT2 or Flamingo, which uses 80B parameters and is pre-trained for both image-language and vision-language tasks.

Our results outperform the state-of-the-art on open-vocabulary detection fine-tuning as is also shown below.

MAMMUT open-vocabulary detection results on the LVIS dataset compared to state-of-the-art methods. We report the average precisions for rare classes (APr) as is previously adopted in the literature.

Key ingredients

We show that joint training of both contrastive and text-generative objectives is not an easy task, and in our ablations we find that these tasks are served better by different design choices. We see that fewer cross-attention connections are better for retrieval tasks, but more are preferred by VQA tasks. Yet, while this shows that our model’s design choices might be suboptimal for individual tasks, our model is more effective than more complex, or larger, models.

Ablation studies showing that fewer cross-attention connections (1-2) are better for retrieval tasks (top), whereas more connections favor text-generative tasks such as VQA (bottom).


We presented MaMMUT, a simple and compact vision-encoder language-decoder model that jointly trains a number of conflicting objectives to reconcile contrastive-like and text-generative tasks. Our model also serves as a foundation for many more vision-language tasks, achieving state-of-the-art or competitive performance on image-text and text-image retrieval, videoQA, video captioning, open-vocabulary detection and VQA. We hope it can be further used for more multimodal applications.


The work described is co-authored by: Weicheng Kuo, AJ Piergiovanni, Dahun Kim, Xiyang Luo, Ben Caine, Wei Li, Abhijit Ogale, Luowei Zhou, Andrew Dai, Zhifeng Chen, Claire Cui, and Anelia Angelova. We would like to thank Mojtaba Seyedhosseini, Vijay Vasudevan, Priya Goyal, Jiahui Yu, Zirui Wang, Yonghui Wu, Runze Li, Jie Mei, Radu Soricut, Qingqing Huang, Andy Ly, Nan Du, Yuxin Wu, Tom Duerig, Paul Natsev, Zoubin Ghahramani for their help and support.

Source: Google AI Blog

Robust and efficient medical imaging with self-supervision

Despite recent progress in the field of medical artificial intelligence (AI), most existing models are narrow, single-task systems that require large quantities of labeled data to train. Moreover, these models cannot be easily reused in new clinical contexts as they often require the collection, de-identification and annotation of site-specific data for every new deployment environment, which is both laborious and expensive. This problem of data-efficient generalization (a model’s ability to generalize to new settings using minimal new data) continues to be a key translational challenge for medical machine learning (ML) models and has in turn, prevented their broad uptake in real world healthcare settings.

The emergence of foundation models offers a significant opportunity to rethink development of medical AI to make it more performant, safer, and equitable. These models are trained using data at scale, often by self-supervised learning. This process results in generalist models that can rapidly be adapted to new tasks and environments with less need for supervised data. With foundation models, it may be possible to safely and efficiently deploy models across various clinical contexts and environments.

In “Robust and Efficient MEDical Imaging with Self-supervision” (REMEDIS), to be published in Nature Biomedical Engineering, we introduce a unified large-scale self-supervised learning framework for building foundation medical imaging models. This strategy combines large scale supervised transfer learning with self-supervised learning and requires minimal task-specific customization. REMEDIS shows significant improvement in data-efficient generalization across medical imaging tasks and modalities with a 3–100x reduction in site-specific data for adapting models to new clinical contexts and environments. Building on this, we are excited to announce Medical AI Research Foundations (hosted by PhysioNet), an expansion of the public release of chest X-ray Foundations in 2022. Medical AI Research Foundations is a collection of open-source non-diagnostic models (starting with REMEDIS models), APIs, and resources to help researchers and developers accelerate medical AI research.

Large scale self-supervision for medical imaging

REMEDIS uses a combination of natural (non-medical) images and unlabeled medical images to develop strong medical imaging foundation models. Its pre-training strategy consists of two steps. The first involves supervised representation learning on a large-scale dataset of labeled natural images (pulled from Imagenet 21k or JFT) using the Big Transfer (BiT) method.

The second step involves intermediate self-supervised learning, which does not require any labels and instead, trains a model to learn medical data representations independently of labels. The specific approach used for pre-training and learning representations is SimCLR. The method works by maximizing agreement between differently augmented views of the same training example via a contrastive loss in a hidden layer of a feed-forward neural network with multilayer perceptron (MLP) outputs. However, REMEDIS is equally compatible with other contrastive self-supervised learning methods. This training method is applicable for healthcare environments as many hospitals acquire raw data (images) as a routine practice. While processes would have to be implemented to make this data usable within models (i.e., patient consent prior to gathering the data, de-identification, etc.), the costly, time-consuming, and difficult task of labeling that data could be avoided using REMEDIS.

REMEDIS leverages large-scale supervised learning using natural images and self-supervised learning using unlabeled medical data to create strong foundation models for medical imaging.

Given ML model parameter constraints, it is important that our proposed approach works when using both small and large model architecture sizes. To study this in detail, we considered two ResNet architectures with commonly used depth and width multipliers, ResNet-50 (1×) and ResNet-152 (2×) as the backbone encoder networks.

After pre-training, the model was fine-tuned using labeled task-specific medical data and evaluated for in-distribution task performance. In addition, to evaluate the data-efficient generalization, the model was also optionally fine-tuned using small amounts of out-of-distribution (OOD) data.

REMEDIS starts with representations initialized using large-scale natural image pretraining following the Big Transfer (BiT) method. We then adapt the model to the medical domain using intermediate contrastive self-supervised learning without using any labeled medical data. Finally, we fine-tune the model to specific downstream medical imaging tasks. We evaluate the ML model both in an in-distribution (ID) setting and in an out-of-distribution (OOD) setting to establish the data-efficient generalization performance of the model.

Evaluation and results

To evaluate the REMEDIS model’s performance, we simulate realistic scenarios using retrospective de-identified data across a broad range of medical imaging tasks and modalities, including dermatology, retinal imaging, chest X-ray interpretation, pathology and mammography. We further introduce the notion of data-efficient generalization, capturing the model’s ability to generalize to new deployment distributions with a significantly reduced need for expert annotated data from the new clinical setting. In-distribution performance is measured as (1) improvement in zero-shot generalization to OOD settings (assessing performance in an OOD evaluation set, with zero access to training data from the OOD dataset) and (2) significant reduction in the need for annotated data from the OOD settings to reach performance equivalent to clinical experts (or threshold demonstrating clinical utility). REMEDIS exhibits significantly improved in-distribution performance with up to 11.5% relative improvement in diagnostic accuracy over a strongly supervised baseline.

More importantly, our strategy leads to data-efficient generalization of medical imaging models, matching strong supervised baselines resulting in a 3–100x reduction in the need for retraining data. While SimCLR is the primary self-supervised learning approach used in the study, we also show that REMEDIS is compatible with other approaches, such as MoCo-V2, RELIC and Barlow Twins. Furthermore, the approach works across model architecture sizes.

REMEDIS outperformed the supervised baseline pre-trained on JFT-300M for various medical tasks and demonstrated improved data-efficient generalization, reducing data needs by 3–100x for adapting models to new clinical settings. This could potentially translate to significant reduction in clinician hours saved annotating data and cost of developing robust medical imaging systems.
REMEDIS is compatible with MoCo-V2, RELIC and Barlow Twins as alternate self-supervised learning strategies. All the REMEDIS variants lead to data-efficient generalization improvements over the strong supervised baseline for dermatology condition classification (T1), diabetic macular edema classification (T2), and chest X-ray condition classification (T3). The gray shaded area indicates the performance of the strong supervised baseline pre-trained on JFT.

Medical AI Research Foundations

Building on REMEDIS, we are excited to announce Medical AI Research Foundations, an expansion of the public release of chest X-ray Foundations in 2022. Medical AI Research Foundations is a repository of open-source medical foundation models hosted by PhysioNet. This expands the previous API-based approach to also encompass non-diagnostic models, to help researchers and developers accelerate their medical AI research. We believe that REMEDIS and the release of the Medical AI Research Foundations are a step toward building medical models that can generalize across healthcare settings and tasks.

We are seeding Medical AI Research Foundations with REMEDIS models for chest X-ray and pathology (with related code). Whereas the existing chest X-ray Foundation approach focuses on providing frozen embeddings for application-specific fine tuning from a model trained on several large private datasets, the REMEDIS models (trained on public datasets) enable users to fine-tune end-to-end for their application, and to run on local devices. We recommend users test different approaches based on their unique needs for their desired application. We expect to add more models and resources for training medical foundation models such as datasets and benchmarks in the future. We also welcome the medical AI research community to contribute to this.


These results suggest that REMEDIS has the potential to significantly accelerate the development of ML systems for medical imaging, which can preserve their strong performance when deployed in a variety of changing contexts. We believe this is an important step forward for medical imaging AI to deliver a broad impact. Beyond the experimental results presented, the approach and insights described here have been integrated into several of Google’s medical imaging research projects, such as dermatology, mammography and radiology among others. We’re using a similar self-supervised learning approach with our non-imaging foundation model efforts, such as Med-PaLM and Med-PaLM 2.

With REMEDIS, we demonstrated the potential of foundation models for medical imaging applications. Such models hold exciting possibilities in medical applications with the opportunity of multimodal representation learning. The practice of medicine is inherently multimodal and incorporates information from images, electronic health records, sensors, wearables, genomics and more. We believe ML systems that leverage these data at scale using self-supervised learning with careful consideration of privacy, safety, fairness and ethics will help lay the groundwork for the next generation of learning health systems that scale world-class healthcare to everyone.


This work involved extensive collaborative efforts from a multidisciplinary team of researchers, software engineers, clinicians, and cross-functional contributors across Google Health AI and Google Brain. In particular, we would like to thank our first co-author Jan Freyberg and our lead senior authors of these projects, Vivek Natarajan, Alan Karthikesalingam, Mohammad Norouzi and Neil Houlsby for their invaluable contributions and support. We also thank Lauren Winer, Sami Lachgar, Yun Liu and Karan Singhal for their feedback on this post and Tom Small for support in creating the visuals. Finally, we also thank the PhysioNet team for their support on hosting Medical AI Research Foundations. Users with questions can reach out to medical-ai-research-foundations at google.com.

Source: Google AI Blog

Scaling vision transformers to 22 billion parameters

Large Language Models (LLMs) like PaLM or GPT-3 showed that scaling transformers to hundreds of billions of parameters improves performance and unlocks emergent abilities. The biggest dense models for image understanding, however, have reached only 4 billion parameters, despite research indicating that promising multimodal models like PaLI continue to benefit from scaling vision models alongside their language counterparts. Motivated by this, and the results from scaling LLMs, we decided to undertake the next step in the journey of scaling the Vision Transformer.

In “Scaling Vision Transformers to 22 Billion Parameters”, we introduce the biggest dense vision model, ViT-22B. It is 5.5x larger than the previous largest vision backbone, ViT-e, which has 4 billion parameters. To enable this scaling, ViT-22B incorporates ideas from scaling text models like PaLM, with improvements to both training stability (using QK normalization) and training efficiency (with a novel approach called asynchronous parallel linear operations). As a result of its modified architecture, efficient sharding recipe, and bespoke implementation, it was able to be trained on Cloud TPUs with a high hardware utilization1. ViT-22B advances the state of the art on many vision tasks using frozen representations, or with full fine-tuning. Further, the model has also been successfully used in PaLM-e, which showed that a large model combining ViT-22B with a language model can significantly advance the state of the art in robotics tasks.


Our work builds on many advances from LLMs, such as PaLM and GPT-3. Compared to the standard Vision Transformer architecture, we use parallel layers, an approach in which attention and MLP blocks are executed in parallel, instead of sequentially as in the standard Transformer. This approach was used in PaLM and reduced training time by 15%.

Secondly, ViT-22B omits biases in the QKV projections, part of the self-attention mechanism, and in the LayerNorms, which increases utilization by 3%. The diagram below shows the modified transformer architecture used in ViT-22B:

ViT-22B transformer encoder architecture uses parallel feed-forward layers, omits biases in QKV and LayerNorm layers and normalizes Query and Key projections.

Models at this scale necessitate “sharding” — distributing the model parameters in different compute devices. Alongside this, we also shard the activations (the intermediate representations of an input). Even something as simple as a matrix multiplication necessitates extra care, as both the input and the matrix itself are distributed across devices. We develop an approach called asynchronous parallel linear operations, whereby communications of activations and weights between devices occur at the same time as computations in the matrix multiply unit (the part of the TPU holding the vast majority of the computational capacity). This asynchronous approach minimizes the time waiting on incoming communication, thus increasing device efficiency. The animation below shows an example computation and communication pattern for a matrix multiplication.

Asynchronized parallel linear operation. The goal is to compute the matrix multiplication y = Ax, but both the matrix A and activation x are distributed across different devices. Here we illustrate how it can be done with overlapping communication and computation across devices. The matrix A is column-sharded across the devices, each holding a contiguous slice, each block represented as Aij. More details are in the paper.

At first, the new model scale resulted in severe training instabilities. The normalization approach of Gilmer et al. (2023, upcoming) resolved these issues, enabling smooth and stable model training; this is illustrated below with example training progressions.

The effect of normalizing the queries and keys (QK normalization) in the self-attention layer on the training dynamics. Without QK normalization (red) gradients become unstable and the training loss diverges.


Here we highlight some results of ViT-22B. Note that in the paper we also explore several other problem domains, like video classification, depth estimation, and semantic segmentation.

To illustrate the richness of the learned representation, we train a text model to produce representations that align text and image representations (using LiT-tuning). Below we show several results for out-of-distribution images generated by Parti and Imagen:

Examples of image+text understanding for ViT-22B paired with a text model. The graph shows normalized probability distribution for each description of an image.

Human object recognition alignment

To find out how aligned ViT-22B classification decisions are with human classification decisions, we evaluated ViT-22B fine-tuned with different resolutions on out-of-distribution (OOD) datasets for which human comparison data is available via the model-vs-human toolbox. This toolbox measures three key metrics: How well do models cope with distortions (accuracy)? How different are human and model accuracies (accuracy difference)? Finally, how similar are human and model error patterns (error consistency)? While not all fine-tuning resolutions perform equally well, ViT-22B variants are state of the art for all three metrics. Furthermore, the ViT-22B models also have the highest ever recorded shape bias in vision models. This means that they mostly use object shape, rather than object texture, to inform classification decisions — a strategy known from human perception (which has a shape bias of 96%). Standard models (e.g., ResNet-50, which has aa ~20–30% shape bias) often classify images like the cat with elephant texture below according to the texture (elephant); models with a high shape bias tend to focus on the shape instead (cat). While there are still many important differences between human and model perception, ViT-22B shows increased similarities to human visual object recognition.

Cat or elephant? Car or clock? Bird or bicycle? Example images with the shape of one object and the texture of a different object, used to measure shape/texture bias.
Shape bias evaluation (higher = more shape-biased). Many vision models have a low shape / high texture bias, whereas ViT-22B fine-tuned on ImageNet (red, green, blue trained on 4B images as indicated by brackets after model names, unless trained on ImageNet only) have the highest shape bias recorded in a ML model to date, bringing them closer to a human-like shape bias.

Out-of-distribution performance

Measuring performance on OOD datasets helps assess generalization. In this experiment we construct label-maps (mappings of labels between datasets) from JFT to ImageNet and also from ImageNet to different out-of-distribution datasets like ObjectNet (results after pre-training on this data shown in the left curve below). Then the models are fully fine-tuned on ImageNet.

We observe that scaling Vision Transformers increases OOD performance: even though ImageNet accuracy saturates, we see a significant increase on ObjectNet from ViT-e to ViT-22B (shown by the three orange dots in the upper right below).

Even though ImageNet accuracy saturates, we see a significant increase in performance on ObjectNet from ViT-e/14 to ViT-22B.

Linear probe

Linear probe is a technique where a single linear layer is trained on top of a frozen model. Compared to full fine-tuning, this is much cheaper to train and easier to set up. We observed that the linear probe of ViT-22B performance approaches that of state-of-the-art full fine-tuning of smaller models using high-resolution images (training with higher resolution is generally much more expensive, but for many tasks it yields better results). Here are results of a linear probe trained on the ImageNet dataset and evaluated on the ImageNet validation dataset and other OOD ImageNet datasets.

Linear probe results trained on ImageNet, evaluated on Imagenet-ReaL, ImageNet-v2, ObjectNet, ImageNet-R and ImageNet-A datasets. High-resolution fine-tuned ViT-e/14 provided as a reference.


The knowledge of the bigger model can be transferred to a smaller model using the distillation method. This is helpful as big models are slower and more expensive to use. We found that ViT-22B knowledge can be transferred to smaller models like ViT-B/16 and ViT-L/16, achieving a new state of the art on ImageNet for those model sizes.

Model Approach (dataset) ImageNet1k Accuracy
ViT-B/16       Transformers for Image Recognition at Scale (JFT)       84.2
Scaling Vision Transformers (JFT) 86.6
DeiT III: Revenge of the ViT (INet21k) 86.7
Distilled from ViT-22B (JFT) 88.6
ViT-L/16 Transformers for Image Recognition at Scale (JFT) 87.1
Scaling Vision Transformers (JFT) 88.5
DeiT III: Revenge of the ViT (INet21k) 87.7
Distilled from ViT-22B (JFT) 89.6

Fairness and bias

ML models can be susceptible to unintended unfair biases, such as picking up spurious correlations (measured using demographic parity) or having performance gaps across subgroups. We show that scaling up the size helps in mitigating such issues.

First, scale offers a more favorable tradeoff frontier — performance improves with scale even when the model is post-processed after training to control its level of demographic parity below a prescribed, tolerable level. Importantly, this holds not only when performance is measured in terms of accuracy, but also other metrics, such as calibration, which is a statistical measure of the truthfulness of the model's estimated probabilities. Second, classification of all subgroups tends to improve with scale as demonstrated below. Third, ViT-22B reduces the performance gap across subgroups.

Top: Accuracy for each subgroup in CelebA before debiasing. Bottom: The y-axis shows the absolute difference in performance across the two specific subgroups highlighted in this example: females and males. ViT-22B has a small gap in performance compared to smaller ViT architectures.


We have presented ViT-22B, currently the largest vision transformer model at 22 billion parameters. With small but critical changes to the original architecture, we achieved excellent hardware utilization and training stability, yielding a model that advances the state of the art on several benchmarks. Great performance can be achieved using the frozen model to produce embeddings and then training thin layers on top. Our evaluations further show that ViT-22B shows increased similarities to human visual perception when it comes to shape and texture bias, and offers benefits in fairness and robustness, when compared to existing models.


This is a joint work of Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaleldin Fathy, Elsayed Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Patrick Collier, Alexey Gritsenko, Vighnesh Birodkar, Cristina Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetić, Dustin Tran, Thomas Kipf, Mario Lučić, Xiaohua Zhai, Daniel Keysers Jeremiah Harmsen, and Neil Houlsby

We would like to thank Jasper Uijlings, Jeremy Cohen, Arushi Goel, Radu Soricut, Xingyi Zhou, Lluis Castrejon, Adam Paszke, Joelle Barral, Federico Lebron, Blake Hechtman, and Peter Hawkins. Their expertise and unwavering support played a crucial role in the completion of this paper. We also acknowledge the collaboration and dedication of the talented researchers and engineers at Google Research.

1Note: ViT-22B has 54.9% model FLOPs utilization (MFU) while PaLM reported 46.2% MFU and we measured 44.0% MFU for ViT-e on the same hardware. 

Source: Google AI Blog

Vid2Seq: a pretrained visual language model for describing multi-event videos

Videos have become an increasingly important part of our daily lives, spanning fields such as entertainment, education, and communication. Understanding the content of videos, however, is a challenging task as videos often contain multiple events occurring at different time scales. For example, a video of a musher hitching up dogs to a dog sled before they all race away involves a long event (the dogs pulling the sled) and a short event (the dogs being hitched to the sled). One way to spur research in video understanding is via the task of dense video captioning, which consists of temporally localizing and describing all events in a minutes-long video. This differs from single image captioning and standard video captioning, which consists of describing short videos with a single sentence.

Dense video captioning systems have wide applications, such as making videos accessible to people with visual or auditory impairments, automatically generating chapters for videos, or improving the search of video moments in large databases. Current dense video captioning approaches, however, have several limitations — for example, they often contain highly specialized task-specific components, which make it challenging to integrate them into powerful foundation models. Furthermore, they are often trained exclusively on manually annotated datasets, which are very difficult to obtain and hence are not a scalable solution.

In this post, we introduce “Vid2Seq: Large-Scale Pretraining of a Visual Language Model for Dense Video Captioning”, to appear at CVPR 2023. The Vid2Seq architecture augments a language model with special time tokens, allowing it to seamlessly predict event boundaries and textual descriptions in the same output sequence. In order to pre-train this unified model, we leverage unlabeled narrated videos by reformulating sentence boundaries of transcribed speech as pseudo-event boundaries, and using the transcribed speech sentences as pseudo-event captions. The resulting Vid2Seq model pre-trained on millions of narrated videos improves the state of the art on a variety of dense video captioning benchmarks including YouCook2, ViTT and ActivityNet Captions. Vid2Seq also generalizes well to the few-shot dense video captioning setting, the video paragraph captioning task, and the standard video captioning task. Finally, we have also released the code for Vid2Seq here.

Vid2Seq is a visual language model that predicts dense event captions together with their temporal grounding in a video by generating a single sequence of tokens.

A visual language model for dense video captioning

Multimodal transformer architectures have improved the state of the art on a wide range of video tasks, such as action recognition. However it is not straightforward to adapt such an architecture to the complex task of jointly localizing and captioning events in minutes-long videos.

For a general overview of how we achieve this, we augment a visual language model with special time tokens (like text tokens) that represent discretized timestamps in the video, similar to Pix2Seq in the spatial domain. Given visual inputs, the resulting Vid2Seq model can both take as input and generate sequences of text and time tokens. First, this enables the Vid2Seq model to understand the temporal information of the transcribed speech input, which is cast as a single sequence of tokens. Second, this allows Vid2Seq to jointly predict dense event captions and temporally ground them in the video while generating a single sequence of tokens.

The Vid2Seq architecture includes a visual encoder and a text encoder, which encode the video frames and the transcribed speech input, respectively. The resulting encodings are then forwarded to a text decoder, which autoregressively predicts the output sequence of dense event captions together with their temporal localization in the video. The architecture is initialized with a powerful visual backbone and a strong language model.

Vid2Seq model overview: We formulate dense event captioning as a sequence-to-sequence problem, using special time tokens to allow the model to seamlessly understand and generate sequences of tokens containing both textual semantic information and temporal localization information grounding each text sentence in the video.

Large-scale pre-training on untrimmed narrated videos

Due to the dense nature of the task, the manual collection of annotations for dense video captioning is particularly expensive. Hence we pre-train the Vid2Seq model using unlabeled narrated videos, which are easily available at scale. In particular, we use the YT-Temporal-1B dataset, which includes 18 million narrated videos covering a wide range of domains.

We use transcribed speech sentences and their corresponding timestamps as supervision, which are cast as a single sequence of tokens. We pre-train Vid2Seq with a generative objective that teaches the decoder to predict the transcribed speech sequence given visual inputs only, and a denoising objective that encourages multimodal learning by requiring the model to predict masked tokens given a noisy transcribed speech sequence and visual inputs. In particular, noise is added to the speech sequence by randomly masking out spans of tokens.

Vid2Seq is pre-trained on unlabeled narrated videos with a generative objective (top) and a denoising objective (bottom).

Results on downstream dense video captioning benchmarks

The resulting pre-trained Vid2Seq model can be fine-tuned on downstream tasks with a simple maximum likelihood objective using teacher forcing (i.e., predicting the next token given previous ground-truth tokens). After fine-tuning, Vid2Seq notably improves the state of the art on three standard downstream dense video captioning benchmarks (ActivityNet Captions, YouCook2 and ViTT) and two video clip captioning benchmarks (MSR-VTT, MSVD). In our paper we provide additional ablation studies, qualitative results, as well as results in the few-shot settings and in the video paragraph captioning task.

Comparison to state-of-the-art methods for dense video captioning (left) and for video clip captioning (right), on the CIDEr metric (higher is better).


We introduce Vid2Seq, a novel visual language model for dense video captioning that simply predicts all event boundaries and captions as a single sequence of tokens. Vid2Seq can be effectively pretrained on unlabeled narrated videos at scale, and achieves state-of-the-art results on various downstream dense video captioning benchmarks. Learn more from the paper and grab the code here.


This research was conducted by Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic and Cordelia Schmid.

Source: Google AI Blog

PaLM-E: An embodied multimodal language model

Recent years have seen tremendous advances across machine learning domains, from models that can explain jokes or answer visual questions in a variety of languages to those that can produce images based on text descriptions. Such innovations have been possible due to the increase in availability of large scale datasets along with novel advances that enable the training of models on these data. While scaling of robotics models has seen some success, it is outpaced by other domains due to a lack of datasets available on a scale comparable to large text corpora or image datasets.

Today we introduce PaLM-E, a new generalist robotics model that overcomes these issues by transferring knowledge from varied visual and language domains to a robotics system. We began with PaLM, a powerful large language model, and “embodied” it (the “E” in PaLM-E), by complementing it with sensor data from the robotic agent. This is the key difference from prior efforts to bring large language models to robotics — rather than relying on only textual input, with PaLM-E we train the language model to directly ingest raw streams of robot sensor data. The resulting model not only enables highly effective robot learning, but is also a state-of-the-art general-purpose visual-language model, while maintaining excellent language-only task capabilities.

An embodied  language model, and also a visual-language generalist

On the one hand, PaLM-E was primarily developed to be a model for robotics, and it solves a variety of tasks on multiple types of robots and for multiple modalities (images, robot states, and neural scene representations). At the same time, PaLM-E is a generally-capable vision-and-language model. It can perform visual tasks, such as describing images, detecting objects, or classifying scenes, and is also proficient at language tasks, like quoting poetry, solving math equations or generating code.

PaLM-E combines our most recent large language model, PaLM, together with one of our most advanced vision models, ViT-22B. The largest instantiation of this approach, built on PaLM-540B, is called PaLM-E-562B and sets a new state of the art on the visual-language OK-VQA benchmark, without task-specific fine-tuning, and while retaining essentially the same general language performance as PaLM-540B.

How does PaLM-E work?

Technically, PaLM-E works by injecting observations into a pre-trained language model. This is realized by transforming sensor data, e.g., images, into a representation through a procedure that is comparable to how words of natural language are processed by a language model.

Language models rely on a mechanism to represent text mathematically in a way that neural networks can process. This is achieved by first splitting the text into so-called tokens that encode (sub)words, each of which is associated with a high-dimensional vector of numbers, the token embedding. The language model is then able to apply mathematical operations (e.g., matrix multiplication) on the resulting sequence of vectors to predict the next, most likely word token. By feeding the newly predicted word back to the input, the language model can iteratively generate a longer and longer text.

The inputs to PaLM-E are text and other modalities — images, robot states, scene embeddings, etc. — in an arbitrary order, which we call "multimodal sentences". For example, an input might look like, "What happened between <img_1> and <img_2>?", where <img_1> and <img_2> are two images. The output is text generated auto-regressively by PaLM-E, which could be an answer to a question, or a sequence of decisions in text form.

PaLM-E model architecture, showing how PaLM-E ingests different modalities (states and/or images) and addresses tasks through multimodal language modeling.

The idea of PaLM-E is to train encoders that convert a variety of inputs into the same space as the natural word token embeddings. These continuous inputs are mapped into something that resembles "words" (although they do not necessarily form discrete sets). Since both the word and image embeddings now have the same dimensionality, they can be fed into the language model.

We initialize PaLM-E for training with pre-trained models for both the language (PaLM) and vision components (Vision Transformer, a.k.a. ViT). All parameters of the model can be updated during training.

Transferring knowledge from large-scale training to robots

PaLM-E offers a new paradigm for training a generalist model, which is achieved by framing robot tasks and vision-language tasks together through a common representation: taking images and text as input, and outputting text. A key result is that PaLM-E attains significant positive knowledge transfer from both the vision and language domains, improving the effectiveness of robot learning.

Positive transfer of knowledge from general vision-language tasks results in more effective robot learning, shown for three different robot embodiments and domains.

Results show that PaLM-E can address a large set of robotics, vision and language tasks simultaneously without performance degradation compared to training individual models on individual tasks. Further, the visual-language data actually significantly improves the performance of the robot tasks. This transfer enables PaLM-E to learn robotics tasks efficiently in terms of the number of examples it requires to solve a task.


We evaluate PaLM-E on three robotic environments, two of which involve real robots, as well as general vision-language tasks such as visual question answering (VQA), image captioning, and general language tasks. When PaLM-E is tasked with making decisions on a robot, we pair it with a low-level language-to-action policy to translate text into low-level robot actions.

In the first example below, a person asks a mobile robot to bring a bag of chips to them. To successfully complete the task, PaLM-E produces a plan to find the drawer and open it and then responds to changes in the world by updating its plan as it executes the task. In the second example, the robot is asked to grab a green block. Even though the block has not been seen by that robot, PaLM-E still generates a step-by-step plan that generalizes beyond the training data of that robot.

PaLM-E controls a mobile robot operating in a kitchen environment. Left: The task is to get a chip bag. PaLM-E shows robustness against adversarial disturbances, such as putting the chip bag back into the drawer. Right: The final steps of executing a plan to retrieve a previously unseen block (green star). This capability is facilitated by transfer learning from the vision and language models.

In the second environment below, the same PaLM-E model solves very long-horizon, precise tasks, such as “sort the blocks by colors into corners,” on a different type of robot. It directly looks at the images and produces a sequence of shorter textually-represented actions — e.g., “Push the blue cube to the bottom right corner,” “Push the blue triangle there too.” — long-horizon tasks that were out of scope for autonomous completion, even in our own most recent models. We also demonstrate the ability to generalize to new tasks not seen during training time (zero-shot generalization), such as pushing red blocks to the coffee cup.

PaLM-E controlling a tabletop robot to successfully complete long-horizon tasks.

The third robot environment is inspired by the field of task and motion planning (TAMP), which studies combinatorially challenging planning tasks (rearranging objects) that confront the robot with a very high number of possible action sequences. We show that with a modest amount of training data from an expert TAMP planner, PaLM-E is not only able to also solve these tasks, but it also leverages visual and language knowledge transfer in order to more effectively do so.

PaLM-E produces plans for a task and motion planning environment.

As a visual-language generalist, PaLM-E is a competitive model, even compared with the best vision-language-only models, including Flamingo and PaLI. In particular, PaLM-E-562B achieves the highest number ever reported on the challenging OK-VQA dataset, which requires not only visual understanding but also external knowledge of the world. Further, this result is reached with a generalist model, without fine-tuning specifically on only that task.

PaLM-E exhibits capabilities like visual chain-of-thought reasoning in which the model breaks down its answering process in smaller steps, an ability that has so far only been demonstrated in the language-only domain. The model also demonstrates the ability to perform inference on multiple images although being trained on only single-image prompts. The image of the New York Knicks and Boston Celtics is under the terms CC-by-2.0 and was posted to Flickr by kowarski. The image of Kobe Bryant is in the Public Domain. The other images were taken by us.


PaLM-E pushes the boundaries of how generally-capable models can be trained to simultaneously address vision, language and robotics while also being capable of transferring knowledge from vision and language to the robotics domain. There are additional topics investigated in further detail in the paper, such as how to leverage neural scene representations with PaLM-E and also the extent to which PaLM-E, with greater model scale, experiences less catastrophic forgetting of its language capabilities.

PaLM-E not only provides a path towards building more capable robots that benefit from other data sources, but might also be a key enabler to other broader applications using multimodal learning, including the ability to unify tasks that have so far seemed separate.


This work was done in collaboration across several teams at Google, including the Robotics at Google team and the Brain team, and with TU Berlin. Co-authors: Igor Mordatch, Andy Zeng, Aakanksha Chowdhery, Klaus Greff, Mehdi S. M. Sajjadi, Daniel Duckworth, Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Fei Xia, Brian Ichter, Karol Hausman, Tianhe Yu, Quan Vuong, Yevgen Chebotar, Wenlong Huang, Pierre Sermanet, Sergey Levine, Vincent Vanhoucke, and Marc Toussiant. Danny is a PhD student advised by Marc Toussaint at TU Berlin. We also would like to thank several other colleagues for their advice and help, including Xi Chen, Etienne Pot, Sebastian Goodman, Maria Attarian, Ted Xiao, Keerthana Gopalakrishnan, Kehang Han, Henryk Michalewski, Neil Houlsby, Basil Mustafa, Justin Gilmer, Yonghui Wu, Erica Moreira, Victor Gomes, Tom Duerig, Mario Lucic, Henning Meyer, and Kendra Byrne.

Source: Google AI Blog

Announcing the ICDAR 2023 Competition on Hierarchical Text Detection and Recognition

The last few decades have witnessed the rapid development of Optical Character Recognition (OCR) technology, which has evolved from an academic benchmark task used in early breakthroughs of deep learning research to tangible products available in consumer devices and to third party developers for daily use. These OCR products digitize and democratize the valuable information that is stored in paper or image-based sources (e.g., books, magazines, newspapers, forms, street signs, restaurant menus) so that they can be indexed, searched, translated, and further processed by state-of-the-art natural language processing techniques.

Research in scene text detection and recognition (or scene text spotting) has been the major driver of this rapid development through adapting OCR to natural images that have more complex backgrounds than document images. These research efforts, however, focus on the detection and recognition of each individual word in images, without understanding how these words compose sentences and articles.

Layout analysis is another relevant line of research that takes a document image and extracts its structure, i.e., title, paragraphs, headings, figures, tables and captions. These layout analysis efforts are parallel to OCR and have been largely developed as independent techniques that are typically evaluated only on document images. As such, the synergy between OCR and layout analysis remains largely under-explored. We believe that OCR and layout analysis are mutually complementary tasks that enable machine learning to interpret text in images and, when combined, could improve the accuracy and efficiency of both tasks.

With this in mind, we announce the Competition on Hierarchical Text Detection and Recognition (the HierText Challenge), hosted as part of the 17th annual International Conference on Document Analysis and Recognition (ICDAR 2023). The competition is hosted on the Robust Reading Competition website, and represents the first major effort to unify OCR and layout analysis. In this competition, we invite researchers from around the world to build systems that can produce hierarchical annotations of text in images using words clustered into lines and paragraphs. We hope this competition will have a significant and long-term impact on image-based text understanding with the goal to consolidate the research efforts across OCR and layout analysis, and create new signals for downstream information processing tasks.

The concept of hierarchical text representation.

Constructing a hierarchical text dataset

In this competition, we use the HierText dataset that we published at CVPR 2022 with our paper "Towards End-to-End Unified Scene Text Detection and Layout Analysis". It’s the first real-image dataset that provides hierarchical annotations of text, containing word, line, and paragraph level annotations. Here, "words" are defined as sequences of textual characters not interrupted by spaces. "Lines" are then interpreted as "space"-separated clusters of "words" that are logically connected in one direction, and aligned in spatial proximity. Finally, "paragraphs" are composed of "lines" that share the same semantic topic and are geometrically coherent.

To build this dataset, we first annotated images from the Open Images dataset using the Google Cloud Platform (GCP) Text Detection API. We filtered through these annotated images, keeping only images rich in text content and layout structure. Then, we worked with our third-party partners to manually correct all transcriptions and to label words, lines and paragraph composition. As a result, we obtained 11,639 transcribed images, split into three subsets: (1) a train set with 8,281 images, (2) a validation set with 1,724 images, and (3) a test set with 1,634 images. As detailed in the paper, we also checked the overlap between our dataset, TextOCR, and Intel OCR (both of which also extracted annotated images from Open Images), making sure that the test images in the HierText dataset were not also included in the TextOCR or Intel OCR training and validation splits and vice versa. Below, we visualize examples using the HierText dataset and demonstrate the concept of hierarchical text by shading each text entity with different colors. We can see that HierText has a diversity of image domain, text layout, and high text density.

Samples from the HierText dataset. Left: Illustration of each word entity. Middle: Illustration of line clustering. Right: Illustration paragraph clustering.

Dataset with highest density of text

In addition to the novel hierarchical representation, HierText represents a new domain of text images. We note that HierText is currently the most dense publicly available OCR dataset. Below we summarize the characteristics of HierText in comparison with other OCR datasets. HierText identifies 103.8 words per image on average, which is more than 3x the density of TextOCR and 25x more dense than ICDAR-2015. This high density poses unique challenges for detection and recognition, and as a consequence HierText is used as one of the primary datasets for OCR research at Google.

Dataset       Training split       Validation split       Testing split       Words per image      
ICDAR-2015       1,000       0       500       4.4      
TextOCR       21,778       3,124       3,232       32.1      
Intel OCR       19,1059       16,731       0       10.0      
HierText       8,281       1,724       1,634       103.8

Comparing several OCR datasets to the HierText dataset.

Spatial distribution

We also find that text in the HierText dataset has a much more even spatial distribution than other OCR datasets, including TextOCR, Intel OCR, IC19 MLT, COCO-Text and IC19 LSVT. These previous datasets tend to have well-composed images, where text is placed in the middle of the images, and are thus easier to identify. On the contrary, text entities in HierText are broadly distributed across the images. It's proof that our images are from more diverse domains. This characteristic makes HierText uniquely challenging among public OCR datasets.

Spatial distribution of text instances in different datasets.

The HierText challenge

The HierText Challenge represents a novel task and with unique challenges for OCR models. We invite researchers to participate in this challenge and join us in ICDAR 2023 this year in San Jose, CA. We hope this competition will spark research community interest in OCR models with rich information representations that are useful for novel down-stream tasks.


The core contributors to this project are Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii and Michalis Raptis. Ashok Popat and Jake Walker provided valuable advice. We also thank Dimosthenis Karatzas and Sergi Robles from Autonomous University of Barcelona for helping us set up the competition website.

Source: Google AI Blog

A vision-language approach for foundational UI understanding

The computational understanding of user interfaces (UI) is a key step towards achieving intelligent UI behaviors. Previously, we investigated various UI modeling tasks, including widget captioning, screen summarization, and command grounding, that address diverse interaction scenarios such as automation and accessibility. We also demonstrated how machine learning can help user experience practitioners improve UI quality by diagnosing tappability confusion and providing insights for improving UI design. These works along with those developed by others in the field have showcased how deep neural networks can potentially transform end user experiences and the interaction design practice.

With these successes in addressing individual UI tasks, a natural question is whether we can obtain foundational understandings of UIs that can benefit specific UI tasks. As our first attempt to answer this question, we developed a multi-task model to address a range of UI tasks simultaneously. Although the work made some progress, a few challenges remain. Previous UI models heavily rely on UI view hierarchies — i.e., the structure or metadata of a mobile UI screen like the Document Object Model for a webpage — that allow a model to directly acquire detailed information of UI objects on the screen (e.g., their types, text content and positions). This metadata has given previous models advantages over their vision-only counterparts. However, view hierarchies are not always accessible, and are often corrupted with missing object descriptions or misaligned structure information. As a result, despite the short-term gains from using view hierarchies, it may ultimately hamper the model performance and applicability. In addition, previous models had to deal with heterogeneous information across datasets and UI tasks, which often resulted in complex model architectures that were difficult to scale or generalize across tasks.

In “Spotlight: Mobile UI Understanding using Vision-Language Models with a Focus”, accepted for publication at ICLR 2023, we present a vision-only approach that aims to achieve general UI understanding completely from raw pixels. We introduce a unified approach to represent diverse UI tasks, the information for which can be universally represented by two core modalities: vision and language. The vision modality captures what a person would see from a UI screen, and the language modality can be natural language or any token sequences related to the task. We demonstrate that Spotlight substantially improves accuracy on a range of UI tasks, including widget captioning, screen summarization, command grounding and tappability prediction.

Spotlight Model

The Spotlight model input includes a tuple of three items: the screenshot, the region of interest on the screen, and the text description of the task. The output is a text description or response about the region of interest. This simple input and output representation of the model is expressive to capture various UI tasks and allows scalable model architectures. This model design allows a spectrum of learning strategies and setups, from task-specific fine-tuning, to multi-task learning and to few-shot learning. The Spotlight model, as illustrated in the above figure, leverages existing architecture building blocks such as ViT and T5 that are pre-trained in the high-resourced, general vision-language domain, which allows us to build on top of the success of these general domain models.

Because UI tasks are often concerned with a specific object or area on the screen, which requires a model to be able to focus on the object or area of interest, we introduce a Focus Region Extractor to a vision-language model that enables the model to concentrate on the region in light of the screen context.

In particular, we design a Region Summarizer that acquires a latent representation of a screen region based on ViT encodings by using attention queries generated from the bounding box of the region (see paper for more details). Specifically, each coordinate (a scalar value, i.e., the left, top, right or bottom) of the bounding box, denoted as a yellow box on the screenshot, is first embedded via a multilayer perceptron (MLP) as a collection of dense vectors, and then fed to a Transformer model along their coordinate-type embedding. The dense vectors and their corresponding coordinate-type embeddings are color coded to indicate their affiliation with each coordinate value. Coordinate queries then attend to screen encodings output by ViT via cross attention, and the final attention output of the Transformer is used as the region representation for the downstream decoding by T5.

A target region on the screen is summarized by using its bounding box to query into screen encodings from ViT via attentional mechanisms.


We pre-train the Spotlight model using two unlabeled datasets (an internal dataset based on C4 corpus and an internal mobile dataset) with 2.5 million mobile UI screens and 80 million web pages. We then separately fine-tune the pre-trained model for each of the four downstream tasks (captioning, summarization, grounding, and tappability). For widget captioning and screen summarization tasks, we report CIDEr scores, which measure how similar a model text description is to a set of references created by human raters. For command grounding, we report accuracy that measures the percentage of times the model successfully locates a target object in response to a user command. For tappability prediction, we report F1 scores that measure the model’s ability to tell tappable objects from untappable ones.

In this experiment, we compare Spotlight with several benchmark models. Widget Caption uses view hierarchy and the image of each UI object to generate a text description for the object. Similarly, Screen2Words uses view hierarchy and the screenshot as well as auxiliary features (e.g., app description) to generate a summary for the screen. In the same vein, VUT combines screenshots and view hierarchies for performing multiple tasks. Finally, the original Tappability model leverages object metadata from view hierarchy and the screenshot to predict object tappability. Taperception, a follow-up model of Tappability, uses a vision-only tappability prediction approach. We examine two Spotlight model variants with respect to the size of its ViT building block, including B/16 and L/16. Spotlight drastically exceeded the state-of-the-art across four UI modeling tasks.

Widget Caption      97      -      -      -     
Screen2Words      -      61.3      -      -     
VUT      99.3      65.6      82.1      -     
Taperception      -      -      -      85.5     
Tappability      -      -      -      87.9     


L/16      141.8      106.7      95.8      88.4     

We then pursue a more challenging setup where we ask the model to learn multiple tasks simultaneously because a multi-task model can substantially reduce model footprint. As shown in the table below, the experiments showed that our model still performs competitively.

Model      Captioning      Summarization      Grounding      Tappability
VUT multi-task      99.3      65.1      80.8      -     
Spotlight B/16      140      102.7      90.8      89.4     
Spotlight L/16      141.3      99.2      94.2      89.5     

To understand how the Region Summarizer enables Spotlight to focus on a target region and relevant areas on the screen, we analyze the attention weights (which indicate where the model attention is on the screenshot) for both widget captioning and screen summarization tasks. In the figure below, for the widget captioning task, the model predicts “select Chelsea team” for the checkbox on the left side, highlighted with a red bounding box. We can see from its attention heatmap (which illustrates the distribution of attention weights) on the right that the model learns to attend to not only the target region of the check box, but also the text “Chelsea" on the far left to generate the caption. For the screen summarization example, the model predicts “page displaying the tutorial of a learning app” given the screenshot on the left. In this example, the target region is the entire screen, and the model learns to attend to important parts on the screen for summarization.

For the widget captioning task, the attention heatmap shows the model attending to the checkbox, i.e., the target object, and the text label on its left when generating a caption for the object. The red bounding box in the figure is for illustration purposes.
For the screen summarization task that the target region encloses the entire screen, the attention heatmap shows the model attending to various locations on the screen that contribute to generating the summary.


We demonstrate that Spotlight outperforms previous methods that use both screenshots and view hierarchies as the input, and establishes state-of-the-art results on multiple representative UI tasks. These tasks range from accessibility, automation to interaction design and evaluation. Our vision-only approach for mobile UI understanding alleviates the need to use view hierarchy, allows the architecture to easily scale and benefits from the success of large vision-language models pre-trained for the general domain. Compared to recent large vision-language model efforts such as Flamingo and PaLI, Spotlight is relatively small and our experiments show the trend that larger models yield better performance. Spotlight can be easily applied to more UI tasks and potentially advance the fronts of many interaction and user experience tasks.


We thank Mandar Joshi and Tao Li for their help in processing the web pre-training dataset, and Chin-Yi Cheng and Forrest Huang for their feedback for proofreading the paper. Thanks to Tom Small for his help in creating animated figures in this post.

Source: Google AI Blog