Tag Archives: ICLR

Best of both worlds: Achieving scalability and quality in text clustering

Clustering is a fundamental, ubiquitous problem in data mining and unsupervised machine learning, where the goal is to group together similar items. The standard forms of clustering are metric clustering and graph clustering. In metric clustering, a given metric space defines distances between data points, which are grouped together based on their separation. In graph clustering, a given graph connects similar data points through edges, and the clustering process groups data points together based on the connections between them. Both clustering forms are particularly useful for large corpora where class labels can’t be defined. Examples of such corpora are the ever-growing digital text collections of various internet platforms, with applications including organizing and searching documents, identifying patterns in text, and recommending relevant documents to users (see more examples in the following posts: clustering related queries based on user intent and practical differentially private clustering).

The choice of text clustering method often presents a dilemma. One approach is to use embedding models, such as BERT or RoBERTa, to define a metric clustering problem. Another is to utilize cross-attention (CA) models, such as PaLM or GPT, to define a graph clustering problem. CA models can provide highly accurate similarity scores, but constructing the input graph may require a prohibitive quadratic number of inference calls to the model. On the other hand, a metric space can efficiently be defined by distances of embeddings produced by embedding models. However, these similarity distances are typically of substantial lower-quality compared to the similarity signals of CA models, and hence the produced clustering can be of much lower-quality.

An overview of the embedding-based and cross-attention–based similarity scoring functions and their scalability vs. quality dilemma.

Motivated by this, in “KwikBucks: Correlation Clustering with Cheap-Weak and Expensive-Strong Signals”, presented at ICLR 2023, we describe a novel clustering algorithm that effectively combines the scalability benefits from embedding models and the quality from CA models. This graph clustering algorithm has query access to both the CA model and the embedding model, however, we apply a budget on the number of queries made to the CA model. This algorithm uses the CA model to answer edge queries, and benefits from unlimited access to similarity scores from the embedding model. We describe how this proposed setting bridges algorithm design and practical considerations, and can be applied to other clustering problems with similar available scoring functions, such as clustering problems on images and media. We demonstrate how this algorithm yields high-quality clusters with almost a linear number of query calls to the CA model. We have also open-sourced the data used in our experiments.

The clustering algorithm

The KwikBucks algorithm is an extension of the well-known KwikCluster algorithm (Pivot algorithm). The high-level idea is to first select a set of documents (i.e., centers) with no similarity edge between them, and then form clusters around these centers. To obtain the quality from CA models and the runtime efficiency from embedding models, we introduce the novel combo similarity oracle mechanism. In this approach, we utilize the embedding model to guide the selection of queries to be sent to the CA model. When given a set of center documents and a target document, the combo similarity oracle mechanism outputs a center from the set that is similar to the target document, if present. The combo similarity oracle enables us to save on budget by limiting the number of query calls to the CA model when selecting centers and forming clusters. It does this by first ranking centers based on their embedding similarity to the target document, and then querying the CA model for the pair (i.e., target document and ranked center), as shown below.

A combo similarity oracle that for a set of documents and a target document, returns a similar document from the set, if present.

We then perform a post processing step to merge clusters if there is a strong connection between two of them, i.e., when the number of connecting edges is higher than the number of missing edges between two clusters. Additionally, we apply the following steps for further computational savings on queries made to the CA model, and to improve performance at runtime:

  1. We leverage query-efficient correlation clustering to form a set of centers from a set of randomly selected documents instead of selecting these centers from all the documents (in the illustration below, the center nodes are red).
  2. We apply the combo similarity oracle mechanism to perform the cluster assignment step in parallel for all non-center documents and leave documents with no similar center as singletons. In the illustration below, the assignments are depicted by blue arrows and initially two (non-center) nodes are left as singletons due to no assignment.
  3. In the post-processing step, to ensure scalability, we use the embedding similarity scores to filter down the potential mergers (in the illustration below, the green dashed boundaries show these merged clusters).

Illustration of progress of the clustering algorithm on a given graph instance.


We evaluate the novel clustering algorithm on various datasets with different properties using different embedding-based and cross-attention–based models. We compare the clustering algorithm’s performance with the two best performing baselines (see the paper for more details):

To evaluate the quality of clustering, we use precision and recall. Precision is used to calculate the percentage of similar pairs out of all co-clustered pairs and recall is the percentage of co-clustered similar pairs out of all similar pairs. To measure the quality of the obtained solutions from our experiments, we use the F1-score, which is the harmonic mean of the precision and recall, where 1.0 is the highest possible value that indicates perfect precision and recall, and 0 is the lowest possible value that indicates if either precision or recall are zero. The table below reports the F1-score for Kwikbucks and various baselines in the case that we allow only a linear number of queries to the CA model. We show that Kwikbucks offers a substantial boost in performance with a 45% relative improvement compared to the best baseline when averaging across all datasets.

Comparing the clustering algorithm to two baseline algorithms using various public datasets: (1) The query-efficient correlation clustering algorithm for budgeted clustering with access to CA only, and (2) spectral clustering on the k-nearest neighbor (kNN) graph formed by querying the CA model for the k-nearest neighbors of each vertex from embedding-based similarity. Pre-processed datasets can be downloaded here.

The figure below compares the clustering algorithm’s performance with baselines using different query budgets. We observe that KwikBucks consistently outperforms other baselines at various budgets.

A comparison of KwikBucks with top-2 baselines when allowed different budgets for querying the cross-attention model.


Text clustering often presents a dilemma in the choice of similarity function: embedding models are scalable but lack quality, while cross-attention models offer quality but substantially hurt scalability. We present a clustering algorithm that offers the best of both worlds: the scalability of embedding models and the quality of cross-attention models. KwikBucks can also be applied to other clustering problems with multiple similarity oracles of varying accuracy levels. This is validated with an exhaustive set of experiments on various datasets with diverse properties. See the paper for more details.


This project was initiated during Sandeep Silwal’s summer internship at Google in 2022. We would like to express our gratitude to our co-authors, Andrew McCallum, Andrew Nystrom, Deepak Ramachandran, and Sandeep Silwal, for their valuable contributions to this work. We also thank Ravi Kumar and John Guilyard for assistance with this blog post.

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

Google at ICLR 2023

The Eleventh International Conference on Learning Representations (ICLR 2023) is being held this week as a hybrid event in Kigali, Rwanda. We are proud to be a Diamond Sponsor of ICLR 2023, a premier conference on deep learning, where Google researchers contribute at all levels. This year we are presenting over 100 papers and are actively involved in organizing and hosting a number of different events, including workshops and interactive sessions.

If you’re registered for ICLR 2023, we hope you’ll visit the Google booth to learn more about the exciting work we’re doing across topics spanning representation and reinforcement learning, theory and optimization, social impact, safety and privacy, and applications from generative AI to speech and robotics. Continue below to find the many ways in which Google researchers are engaged at ICLR 2023, including workshops, papers, posters and talks (Google affiliations in bold).

Board and Organizing Committee

Board Members include: Shakir Mohamed, Tara Sainath

Senior Program Chairs include: Been Kim

Workshop Chairs include: Aisha Walcott-Bryant, Rose Yu

Diversity, Equity & Inclusion Chairs include: Rosanne Liu

Outstanding Paper awards

Emergence of Maps in the Memories of Blind Navigation Agents
Erik Wijmans, Manolis Savva, Irfan Essa, Stefan Lee, Ari S. Morcos, Dhruv Batra

DreamFusion: Text-to-3D Using 2D Diffusion
Ben Poole, Ajay Jain, Jonathan T. Barron, Ben Mildenhall

Keynote speaker

Learned Optimizers: Why They're the Future, Why They’re Hard, and What They Can Do Now
Jascha Sohl-Dickstein


Kaggle@ICLR 2023: ML Solutions in Africa
Organizers include: Julia Elliott, Phil Culliton, Ray Harvey
Facilitators: Julia Elliot, Walter Reade

Reincarnating Reinforcement Learning (Reincarnating RL)
Organizers include: Rishabh Agarwal, Ted Xiao, Max Schwarzer
Speakers include: Sergey Levine
Panelists include: Marc G. Bellemare, Sergey Levine

Trustworthy and Reliable Large-Scale Machine Learning Models
Organizers include: Sanmi Koyejo
Speakers include: Nicholas Carlini

Physics for Machine Learning (Physics4ML)
Speakers include: Yasaman Bahri

AI for Agent-Based Modelling Community (AI4ABM)
Organizers include: Pablo Samuel Castro

Mathematical and Empirical Understanding of Foundation Models (ME-FoMo)
Organizers include: Mathilde Caron, Tengyu Ma, Hanie Sedghi
Speakers include: Yasaman Bahri, Yann Dauphin

Neurosymbolic Generative Models 2023 (NeSy-GeMs)
Organizers include: Kevin Ellis
Speakers include: Daniel Tarlow, Tuan Anh Le

What Do We Need for Successful Domain Generalization?
Panelists include: Boqing Gong

The 4th Workshop on Practical ML for Developing Countries: Learning Under Limited/Low Resource Settings
Keynote Speaker: Adji Bousso Dieng

Machine Learning for Remote Sensing
Speakers include: Abigail Annkah

Multimodal Representation Learning (MRL): Perks and Pitfalls
Organizers include: Petra Poklukar
Speakers include: Arsha Nagrani

Pitfalls of Limited Data and Computation for Trustworthy ML
Organizers include: Prateek Jain
Speakers include: Nicholas Carlini, Praneeth Netrapalli

Sparsity in Neural Networks: On Practical Limitations and Tradeoffs Between Sustainability and Efficiency
Organizers include: Trevor Gale, Utku Evci
Speakers include: Aakanksha Chowdhery, Jeff Dean

Time Series Representation Learning for Health
Speakers include: Katherine Heller

Deep Learning for Code (DL4C)
Organizers include: Gabriel Orlanski
Speakers include: Alex Polozov, Daniel Tarlow

Affinity Workshops

Tiny Papers Showcase Day (a DEI initiative)
Organizers include: Rosanne Liu


Evolve Smoothly, Fit Consistently: Learning Smooth Latent Dynamics for Advection-Dominated Systems
Zhong Yi Wan, Leonardo Zepeda-Nunez, Anudhyan Boral, Fei Sha

Quantifying Memorization Across Neural Language Models
Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, Chiyuan Zhang

Emergence of Maps in the Memories of Blind Navigation Agents (Outstanding Paper Award)
Erik Wijmans, Manolis Savva, Irfan Essa, Stefan Lee, Ari S. Morcos, Dhruv Batra

Offline Q-Learning on Diverse Multi-task Data Both Scales and Generalizes (see blog post)
Aviral Kumar, Rishabh Agarwal, Xingyang Geng, George Tucker, Sergey Levine

ReAct: Synergizing Reasoning and Acting in Language Models (see blog post)
Shunyu Yao*, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, Yuan Cao

Prompt-to-Prompt Image Editing with Cross-Attention Control
Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, Daniel Cohen-Or

DreamFusion: Text-to-3D Using 2D Diffusion (Outstanding Paper Award)
Ben Poole, Ajay Jain, Jonathan T. Barron, Ben Mildenhall

A System for Morphology-Task Generalization via Unified Representation and Behavior Distillation
Hiroki Furuta, Yusuke Iwasawa, Yutaka Matsuo, Shixiang Shane Gu

Sample-Efficient Reinforcement Learning by Breaking the Replay Ratio Barrier
Pierluca D'Oro, Max Schwarzer, Evgenii Nikishin, Pierre-Luc Bacon, Marc G Bellemare, Aaron Courville

Dichotomy of Control: Separating What You Can Control from What You Cannot
Sherry Yang, Dale Schuurmans, Pieter Abbeel, Ofir Nachum

Fast and Precise: Adjusting Planning Horizon with Adaptive Subgoal Search
Michał Zawalski, Michał Tyrolski, Konrad Czechowski, Tomasz Odrzygóźdź, Damian Stachura, Piotr Piekos, Yuhuai Wu, Łukasz Kucinski, Piotr Miłos

The Trade-Off Between Universality and Label Efficiency of Representations from Contrastive Learning
Zhenmei Shi, Jiefeng Chen, Kunyang Li, Jayaram Raghuram, Xi Wu, Yingyu Liang, Somesh Jha

Sparsity-Constrained Optimal Transport
Tianlin Liu*, Joan Puigcerver, Mathieu Blondel

Unmasking the Lottery Ticket Hypothesis: What's Encoded in a Winning Ticket's Mask?
Mansheej Paul, Feng Chen, Brett W. Larsen, Jonathan Frankle, Surya Ganguli, Gintare Karolina Dziugaite

Extreme Q-Learning: MaxEnt RL without Entropy
Divyansh Garg, Joey Hejna, Matthieu Geist, Stefano Ermon

Draft, Sketch, and Prove: Guiding Formal Theorem Provers with Informal Proofs
Albert Qiaochu Jiang, Sean Welleck, Jin Peng Zhou, Timothee Lacroix, Jiacheng Liu, Wenda Li, Mateja Jamnik, Guillaume Lample, Yuhuai Wu

SimPer: Simple Self-Supervised Learning of Periodic Targets
Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, Daniel McDuff

Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language
Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Marcin Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael S. Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, Pete Florence

What Learning Algorithm Is In-Context Learning? Investigations with Linear Models
Ekin Akyurek*, Dale Schuurmans, Jacob Andreas, Tengyu Ma*, Denny Zhou

Preference Transformer: Modeling Human Preferences Using Transformers for RL
Changyeon Kim, Jongjin Park, Jinwoo Shin, Honglak Lee, Pieter Abbeel, Kimin Lee

Iterative Patch Selection for High-Resolution Image Recognition
Benjamin Bergner, Christoph Lippert, Aravindh Mahendran

Open-Vocabulary Object Detection upon Frozen Vision and Language Models
Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, Anelia Angelova

(Certified!!) Adversarial Robustness for Free!
Nicholas Carlini, Florian Tramér, Krishnamurthy (Dj) Dvijotham, Leslie Rice, Mingjie Sun, J. Zico Kolter

REPAIR: REnormalizing Permuted Activations for Interpolation Repair
Keller Jordan, Hanie Sedghi, Olga Saukh, Rahim Entezari, Behnam Neyshabur

Discrete Predictor-Corrector Diffusion Models for Image Synthesis
José Lezama, Tim Salimans, Lu Jiang, Huiwen Chang, Jonathan Ho, Irfan Essa

Feature Reconstruction From Outputs Can Mitigate Simplicity Bias in Neural Networks
Sravanti Addepalli, Anshul Nasery, Praneeth Netrapalli, Venkatesh Babu R., Prateek Jain

An Exact Poly-time Membership-Queries Algorithm for Extracting a Three-Layer ReLU Network
Amit Daniely, Elad Granot

Language Models Are Multilingual Chain-of-Thought Reasoners
Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, Dipanjan Das, Jason Wei

Scaling Forward Gradient with Local Losses
Mengye Ren*, Simon Kornblith, Renjie Liao, Geoffrey Hinton

Treeformer: Dense Gradient Trees for Efficient Attention Computation
Lovish Madaan, Srinadh Bhojanapalli, Himanshu Jain, Prateek Jain

LilNetX: Lightweight Networks with EXtreme Model Compression and Structured Sparsification
Sharath Girish, Kamal Gupta, Saurabh Singh, Abhinav Shrivastava

DiffusER: Diffusion via Edit-Based Reconstruction
Machel Reid, Vincent J. Hellendoorn, Graham Neubig

Leveraging Unlabeled Data to Track Memorization
Mahsa Forouzesh, Hanie Sedghi, Patrick Thiran

A Mixture-of-Expert Approach to RL-Based Dialogue Management
Yinlam Chow, Aza Tulepbergenov, Ofir Nachum, Dhawal Gupta, Moonkyung Ryu, Mohammad Ghavamzadeh, Craig Boutilier

Easy Differentially Private Linear Regression
Kareem Amin, Matthew Joseph, Monica Ribero, Sergei Vassilvitskii

KwikBucks: Correlation Clustering with Cheap-Weak and Expensive-Strong Signals
Sandeep Silwal*, Sara Ahmadian, Andrew Nystrom, Andrew McCallum, Deepak Ramachandran, Mehran Kazemi

Massively Scaling Heteroscedastic Classifiers
Mark Collier, Rodolphe Jenatton, Basil Mustafa, Neil Houlsby, Jesse Berent, Effrosyni Kokiopoulou

The Lazy Neuron Phenomenon: On Emergence of Activation Sparsity in Transformers
Zonglin Li, Chong You, Srinadh Bhojanapalli, Daliang Li, Ankit Singh Rawat, Sashank J. Reddi, Ke Ye, Felix Chern, Felix Yu, Ruiqi Guo, Sanjiv Kumar

Compositional Semantic Parsing with Large Language Models
Andrew Drozdov, Nathanael Scharli, Ekin Akyurek, Nathan Scales, Xinying Song, Xinyun Chen, Olivier Bousquet, Denny Zhou

Extremely Simple Activation Shaping for Out-of-Distribution Detection
Andrija Djurisic, Nebojsa Bozanic, Arjun Ashok, Rosanne Liu

Long Range Language Modeling via Gated State Spaces
Harsh Mehta, Ankit Gupta, Ashok Cutkosky, Behnam Neyshabur

Investigating Multi-task Pretraining and Generalization in Reinforcement Learning
Adrien Ali Taiga, Rishabh Agarwal, Jesse Farebrother, Aaron Courville, Marc G. Bellemare

Learning Low Dimensional State Spaces with Overparameterized Recurrent Neural Nets
Edo Cohen-Karlik, Itamar Menuhin-Gruman, Raja Giryes, Nadav Cohen, Amir Globerson

Weighted Ensemble Self-Supervised Learning
Yangjun Ruan*, Saurabh Singh, Warren Morningstar, Alexander A. Alemi, Sergey Ioffe, Ian Fischer, Joshua V. Dillon

Calibrating Sequence Likelihood Improves Conditional Language Generation
Yao Zhao, Misha Khalman, Rishabh Joshi, Shashi Narayan, Mohammad Saleh, Peter J. Liu

SMART: Sentences as Basic Units for Text Evaluation
Reinald Kim Amplayo, Peter J. Liu, Yao Zhao, Shashi Narayan

Leveraging Importance Weights in Subset Selection
Gui Citovsky, Giulia DeSalvo, Sanjiv Kumar, Srikumar Ramalingam, Afshin Rostamizadeh, Yunjuan Wang*

Proto-Value Networks: Scaling Representation Learning with Auxiliary Tasks
Jesse Farebrother, Joshua Greaves, Rishabh Agarwal, Charline Le Lan, Ross Goroshin, Pablo Samuel Castro, Marc G. Bellemare

An Extensible Multi-modal Multi-task Object Dataset with Materials
Trevor Standley, Ruohan Gao, Dawn Chen, Jiajun Wu, Silvio Savarese

Measuring Forgetting of Memorized Training Examples
Matthew Jagielski, Om Thakkar, Florian Tramér, Daphne Ippolito, Katherine Lee, Nicholas Carlini, Eric Wallace, Shuang Song, Abhradeep Thakurta, Nicolas Papernot, Chiyuan Zhang

Bidirectional Language Models Are Also Few-Shot Learners
Ajay Patel, Bryan Li, Mohammad Sadegh Rasooli, Noah Constant, Colin Raffel, Chris Callison-Burch

Is Attention All That NeRF Needs?
Mukund Varma T., Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang

Automating Nearest Neighbor Search Configuration with Constrained Optimization
Philip Sun, Ruiqi Guo, Sanjiv Kumar

Static Prediction of Runtime Errors by Learning to Execute Programs with External Resource Descriptions
David Bieber, Rishab Goel, Daniel Zheng, Hugo Larochelle, Daniel Tarlow

Composing Ensembles of Pre-trained Models via Iterative Consensus
Shuang Li, Yilun Du, Joshua B. Tenenbaum, Antonio Torralba, Igor Mordatch

Λ-DARTS: Mitigating Performance Collapse by Harmonizing Operation Selection Among Cells
Sajad Movahedi, Melika Adabinejad, Ayyoob Imani, Arezou Keshavarz, Mostafa Dehghani, Azadeh Shakery, Babak N. Araabi

Blurring Diffusion Models
Emiel Hoogeboom, Tim Salimans

Part-Based Models Improve Adversarial Robustness
Chawin Sitawarin, Kornrapat Pongmala, Yizheng Chen, Nicholas Carlini, David Wagner

Learning in Temporally Structured Environments
Matt Jones, Tyler R. Scott, Mengye Ren, Gamaleldin ElSayed, Katherine Hermann, David Mayo, Michael C. Mozer

SlotFormer: Unsupervised Visual Dynamics Simulation with Object-Centric Models
Ziyi Wu, Nikita Dvornik, Klaus Greff, Thomas Kipf, Animesh Garg

Robust Algorithms on Adaptive Inputs from Bounded Adversaries
Yeshwanth Cherapanamjeri, Sandeep Silwal, David P. Woodruff, Fred Zhang, Qiuyi (Richard) Zhang, Samson Zhou

Agnostic Learning of General ReLU Activation Using Gradient Descent
Pranjal Awasthi, Alex Tang, Aravindan Vijayaraghavan

Analog Bits: Generating Discrete Data Using Diffusion Models with Self-Conditioning
Ting Chen, Ruixiang Zhang, Geoffrey Hinton

Any-Scale Balanced Samplers for Discrete Space
Haoran Sun*, Bo Dai, Charles Sutton, Dale Schuurmans, Hanjun Dai

Augmentation with Projection: Towards an Effective and Efficient Data Augmentation Paradigm for Distillation
Ziqi Wang*, Yuexin Wu, Frederick Liu, Daogao Liu, Le Hou, Hongkun Yu, Jing Li, Heng Ji

Beyond Lipschitz: Sharp Generalization and Excess Risk Bounds for Full-Batch GD
Konstantinos E. Nikolakakis, Farzin Haddadpour, Amin Karbasi, Dionysios S. Kalogerias

Causal Estimation for Text Data with (Apparent) Overlap Violations
Lin Gui, Victor Veitch

Contrastive Learning Can Find an Optimal Basis for Approximately View-Invariant Functions
Daniel D. Johnson, Ayoub El Hanchi, Chris J. Maddison

Differentially Private Adaptive Optimization with Delayed Preconditioners
Tian Li, Manzil Zaheer, Ziyu Liu, Sashank Reddi, Brendan McMahan, Virginia Smith

Distributionally Robust Post-hoc Classifiers Under Prior Shifts
Jiaheng Wei*, Harikrishna Narasimhan, Ehsan Amid, Wen-Sheng Chu, Yang Liu, Abhishek Kumar

Human Alignment of Neural Network Representations
Lukas Muttenthaler, Jonas Dippel, Lorenz Linhardt, Robert A. Vandermeulen, Simon Kornblith

Implicit Bias in Leaky ReLU Networks Trained on High-Dimensional Data
Spencer Frei, Gal Vardi, Peter Bartlett, Nathan Srebro, Wei Hu

Koopman Neural Operator Forecaster for Time-Series with Temporal Distributional Shifts
Rui Wang*, Yihe Dong, Sercan Ö. Arik, Rose Yu

Latent Variable Representation for Reinforcement Learning
Tongzheng Ren, Chenjun Xiao, Tianjun Zhang, Na Li, Zhaoran Wang, Sujay Sanghavi, Dale Schuurmans, Bo Dai

Least-to-Most Prompting Enables Complex Reasoning in Large Language Models
Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, Ed Chi

Mind's Eye: Grounded Language Model Reasoning Through Simulation
Ruibo Liu, Jason Wei, Shixiang Shane Gu, Te-Yen Wu, Soroush Vosoughi, Claire Cui, Denny Zhou, Andrew M. Dai

MOAT: Alternating Mobile Convolution and Attention Brings Strong Vision Models
Chenglin Yang*, Siyuan Qiao, Qihang Yu, Xiaoding Yuan, Yukun Zhu, Alan Yuille, Hartwig Adam, Liang-Chieh Chen

Novel View Synthesis with Diffusion Models
Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, Mohammad Norouzi

On Accelerated Perceptrons and Beyond
Guanghui Wang, Rafael Hanashiro, Etash Guha, Jacob Abernethy

On Compositional Uncertainty Quantification for Seq2seq Graph Parsing
Zi Lin*, Du Phan, Panupong Pasupat, Jeremiah Liu, Jingbo Shang

On the Robustness of Safe Reinforcement Learning Under Observational Perturbations
Zuxin Liu, Zijian Guo, Zhepeng Cen, Huan Zhang, Jie Tan, Bo Li, Ding Zhao

Online Low Rank Matrix Completion
Prateek Jain, Soumyabrata Pal

Out-of-Distribution Detection and Selective Generation for Conditional Language Models
Jie Ren, Jiaming Luo, Yao Zhao, Kundan Krishna*, Mohammad Saleh, Balaji Lakshminarayanan, Peter J. Liu

PaLI: A Jointly-Scaled Multilingual Language-Image Model
Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, Radu Soricut

Phenaki: Variable Length Video Generation from Open Domain Textual Descriptions
Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro*, Julius Kunze*, Dumitru Erhan

Promptagator: Few-Shot Dense Retrieval from 8 Examples
Zhuyun Dai, Vincent Y. Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B. Hall, Ming-Wei Chang

Pushing the Accuracy-Group Robustness Frontier with Introspective Self-Play
Jeremiah Zhe Liu, Krishnamurthy Dj Dvijotham, Jihyeon Lee, Quan Yuan, Balaji Lakshminarayanan, Deepak Ramachandran

Re-Imagen: Retrieval-Augmented Text-to-Image Generator Wenhu Chen, Hexiang Hu, Chitwan Saharia, William W. Cohen

Recitation-Augmented Language Models
Zhiqing Sun, Xuezhi Wang, Yi Tay, Yiming Yang, Denny Zhou

Regression with Label Differential Privacy
Badih Ghazi, Pritish Kamath, Ravi Kumar, Ethan Leeman, Pasin Manurangsi, Avinash Varadarajan, Chiyuan Zhang

Revisiting the Entropy Semiring for Neural Speech Recognition
Oscar Chang, Dongseong Hwang, Olivier Siohan

Robust Active Distillation
Cenk Baykal, Khoa Trinh, Fotis Iliopoulos, Gaurav Menghani, Erik Vee

Score-Based Continuous-Time Discrete Diffusion Models
Haoran Sun*, Lijun Yu, Bo Dai, Dale Schuurmans, Hanjun Dai

Self-Consistency Improves Chain of Thought Reasoning in Language Models
Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, Denny Zhou

Self-Supervision Through Random Segments with Autoregressive Coding (RandSAC)
Tianyu Hua, Yonglong Tian, Sucheng Ren, Michalis Raptis, Hang Zhao, Leonid Sigal

Serving Graph Compression for Graph Neural Networks
Si Si, Felix Yu, Ankit Singh Rawat, Cho-Jui Hsieh, Sanjiv Kumar

Sequential Attention for Feature Selection
Taisuke Yasuda*, MohammadHossein Bateni, Lin Chen, Matthew Fahrbach, Gang Fu, Vahab Mirrokni

Sparse Upcycling: Training Mixture-of-Experts from Dense Checkpoints
Aran Komatsuzaki*, Joan Puigcerver, James Lee-Thorp, Carlos Riquelme, Basil Mustafa, Joshua Ainslie, Yi Tay, Mostafa Dehghani, Neil Houlsby

Spectral Decomposition Representation for Reinforcement Learning
Tongzheng Ren, Tianjun Zhang, Lisa Lee, Joseph Gonzalez, Dale Schuurmans, Bo Dai

Spotlight: Mobile UI Understanding Using Vision-Language Models with a Focus (see blog post)
Gang Li, Yang Li

Supervision Complexity and Its Role in Knowledge Distillation
Hrayr Harutyunyan*, Ankit Singh Rawat, Aditya Krishna Menon, Seungyeon Kim, Sanjiv Kumar

Teacher Guided Training: An Efficient Framework for Knowledge Transfer
Manzil Zaheer, Ankit Singh Rawat, Seungyeon Kim, Chong You, Himanshu Jain, Andreas Veit, Rob Fergus, Sanjiv Kumar

TEMPERA: Test-Time Prompt Editing via Reinforcement Learning
Tianjun Zhang, Xuezhi Wang, Denny Zhou, Dale Schuurmans, Joseph E. Gonzalez

UL2: Unifying Language Learning Paradigms
Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Dara Bahri, Tal Schuster, Steven Zheng, Denny Zhou, Neil Houlsby, Donald Metzler

* Work done while at Google

Source: Google AI Blog

Recent advances in deep long-horizon forecasting

Time-series forecasting is an important research area that is critical to several scientific and industrial applications, like retail supply chain optimization, energy and traffic prediction, and weather forecasting. In retail use cases, for example, it has been observed that improving demand forecasting accuracy can meaningfully reduce inventory costs and increase revenue.

Modern time-series applications can involve forecasting hundreds of thousands of correlated time-series (e.g., demands of different products for a retailer) over long horizons (e.g., a quarter or year away at daily granularity). As such, time-series forecasting models need to satisfy the following key criterias:

  1. Ability to handle auxiliary features or covariates: Most use-cases can benefit tremendously from effectively using covariates, for instance, in retail forecasting, holidays and product specific attributes or promotions can affect demand.
  2. Suitable for different data modalities: It should be able to handle sparse count data, e.g., intermittent demand for a product with low volume of sales while also being able to model robust continuous seasonal patterns in traffic forecasting.

A number of neural network–based solutions have been able to show good performance on benchmarks and also support the above criterion. However, these methods are typically slow to train and can be expensive for inference, especially for longer horizons.

In “Long-term Forecasting with TiDE: Time-series Dense Encoder”, we present an all multilayer perceptron (MLP) encoder-decoder architecture for time-series forecasting that achieves superior performance on long horizon time-series forecasting benchmarks when compared to transformer-based solutions, while being 5–10x faster. Then in “On the benefits of maximum likelihood estimation for Regression and Forecasting”, we demonstrate that using a carefully designed training loss function based on maximum likelihood estimation (MLE) can be effective in handling different data modalities. These two works are complementary and can be applied as a part of the same model. In fact, they will be available soon in Google Cloud AI’s Vertex AutoML Forecasting.

TiDE: A simple MLP architecture for fast and accurate forecasting

Deep learning has shown promise in time-series forecasting, outperforming traditional statistical methods, especially for large multivariate datasets. After the success of transformers in natural language processing (NLP), there have been several works evaluating variants of the Transformer architecture for long horizon (the amount of time into the future) forecasting, such as FEDformer and PatchTST. However, other work has suggested that even linear models can outperform these transformer variants on time-series benchmarks. Nonetheless, simple linear models are not expressive enough to handle auxiliary features (e.g., holiday features and promotions for retail demand forecasting) and non-linear dependencies on the past.

We present a scalable MLP-based encoder-decoder model for fast and accurate multi-step forecasting. Our model encodes the past of a time-series and all available features using an MLP encoder. Subsequently, the encoding is combined with future features using an MLP decoder to yield future predictions. The architecture is illustrated below.

TiDE model architecture for multi-step forecasting.

TiDE is more than 10x faster in training compared to transformer-based baselines while being more accurate on benchmarks. Similar gains can be observed in inference as it only scales linearly with the length of the context (the number of time-steps the model looks back) and the prediction horizon. Below on the left, we show that our model can be 10.6% better than the best transformer-based baseline (PatchTST) on a popular traffic forecasting benchmark, in terms of test mean squared error (MSE). On the right, we show that at the same time our model can have much faster inference latency than PatchTST.

Left: MSE on the test set of a popular traffic forecasting benchmark. Right: inference time of TiDE and PatchTST as a function of the look-back length.

Our research demonstrates that we can take advantage of MLP’s linear computational scaling with look-back and horizon sizes without sacrificing accuracy, while transformers scale quadratically in this situation.

Probabilistic loss functions

In most forecasting applications the end user is interested in popular target metrics like the mean absolute percentage error (MAPE), weighted absolute percentage error (WAPE), etc. In such scenarios, the standard approach is to use the same target metric as the loss function while training. In “On the benefits of maximum likelihood estimation for Regression and Forecasting”, accepted at ICLR, we show that this approach might not always be the best. Instead, we advocate using the maximum likelihood loss for a carefully chosen family of distributions (discussed more below) that can capture inductive biases of the dataset during training. In other words, instead of directly outputting point predictions that minimize the target metric, the forecasting neural network predicts the parameters of a distribution in the chosen family that best explains the target data. At inference time, we can predict the statistic from the learned predictive distribution that minimizes the target metric of interest (e.g., the mean minimizes the MSE target metric while the median minimizes the WAPE). Further, we can also easily obtain uncertainty estimates of our forecasts, i.e., we can provide quantile forecasts by estimating the quantiles of the predictive distribution. In several use cases, accurate quantiles are vital, for instance, in demand forecasting a retailer might want to stock for the 90th percentile to guard against worst-case scenarios and avoid lost revenue.

The choice of the distribution family is crucial in such cases. For example, in the context of sparse count data, we might want to have a distribution family that can put more probability on zero, which is commonly known as zero-inflation. We propose a mixture of different distributions with learned mixture weights that can adapt to different data modalities. In the paper, we show that using a mixture of zero and multiple negative binomial distributions works well in a variety of settings as it can adapt to sparsity, multiple modalities, count data, and data with sub-exponential tails.

A mixture of zero and two negative binomial distributions. The weights of the three components, a1, a2 and a3, can be learned during training.

We use this loss function for training Vertex AutoML models on the M5 forecasting competition dataset and show that this simple change can lead to a 6% gain and outperform other benchmarks in the competition metric, weighted root mean squared scaled error (WRMSSE).

M5 Forecasting WRMSSE
Vertex AutoML 0.639 +/- 0.007
Vertex AutoML with probabilistic loss       0.581 +/- 0.007
DeepAR 0.789 +/- 0.025
FEDFormer 0.804 +/- 0.033


We have shown how TiDE, together with probabilistic loss functions, enables fast and accurate forecasting that automatically adapts to different data distributions and modalities and also provides uncertainty estimates for its predictions. It provides state-of-the-art accuracy among neural network–based solutions at a fraction of the cost of previous transformer-based forecasting architectures, for large-scale enterprise forecasting applications. We hope this work will also spur interest in revisiting (both theoretically and empirically) MLP-based deep time-series forecasting models.


This work is the result of a collaboration between several individuals across Google Research and Google Cloud, including (in alphabetical order): Pranjal Awasthi, Dawei Jia, Weihao Kong, Andrew Leach, Shaan Mathur, Petros Mol, Shuxin Nie, Ananda Theertha Suresh, and Rose Yu.

Source: Google AI Blog

Enabling Creative Expression with Concept Activation Vectors

Advances in computer vision and natural language processing continue to unlock new ways of exploring billions of images available on public and searchable websites. Today’s visual search tools make it possible to search with your camera, voice, text, images, or multiple modalities at the same time. However, it remains difficult to input subjective concepts, such as visual tones or moods, into current systems. For this reason, we have been working collaboratively with artists, photographers, and image researchers to explore how machine learning (ML) might enable people to use expressive queries as a way of visually exploring datasets.

Today, we are introducing Mood Board Search, a new ML-powered research tool that uses mood boards as a query over image collections. This enables people to define and evoke visual concepts on their own terms. Mood Board Search can be useful for subjective queries, such as “peaceful”, or for words and individual images that may not be specific enough to produce useful results in a standard search, such as “abstract details in overlooked scenes” or “vibrant color palette that feels part memory, part dream". We developed, and will continue to develop, this research tool in alignment with our AI Principles.

Search Using Mood Boards
With Mood Board Search, our goal is to design a flexible and approachable interface so people without ML expertise can train a computer to recognize a visual concept as they see it. The tool interface is inspired by mood boards, commonly used by people in creative fields to communicate the “feel” of an idea using collections of visual materials.

With Mood Board Search, users can train a computer to recognize visual concepts in image collections.

To get started, simply drag and drop a small number of images that represent the idea you want to convey. Mood Board Search returns the best results when the images share a consistent visual quality, so results are more likely to be relevant with mood boards that share visual similarities in color, pattern, texture, or composition.

It’s also possible to signal which images are more important to a visual concept by upweighting or downweighting images, or by adding images that are the opposite of the concept. Then, users can review and inspect search results to understand which part of an image best matches the visual concept. Focus mode does this by revealing a bounding box around part of the image, while AI crop cuts in directly, making it easier to draw attention to new compositions.

Supported interactions, like AI crop, allow users to see which part of an image best matches their visual concept.

Powered by Concept Activation Vectors (CAVs)
Mood Board Search takes advantage of pre-trained computer vision models, such as GoogLeNet and MobileNet, and a machine learning approach called Concept Activation Vectors (CAVs).

CAVs are a way for machines to represent images (what we understand) using numbers or directions in a neural net’s embedding space (which can be thought of as what machines understand). CAVs can be used as part of a technique, Testing with CAVs (TCAV), to quantify the degree to which a user-defined concept is important to a classification result; e.g., how sensitive a prediction of "zebra" is to the presence of stripes. This is a research approach we open-sourced in 2018, and the work has since been widely applied to medical applications and science to build ML applications that can provide better explanations for what machines see. You can learn more about embedding vectors in general in this Google AI blog post, and our approach to working with TCAVs in Been Kim’s Keynote at ICLR.

In Mood Board Search, we use CAVs to find a model's sensitivity to a mood board created by the user. In other words, each mood board creates a CAV — a direction in embedding space — and the tool searches an image dataset, surfacing images that are the closest match to the CAV. However, the tool takes it one step further, by segmenting each image in the dataset in 15 different ways, to uncover as many relevant compositions as possible. This is the approach behind features like Focus mode and AI crop.

Three artists created visual concepts to share their way of seeing, shown here in an experimental app by design invention studio, Nord Projects.

Because embedding vectors can be learned and re-used across models, tools like Mood Board Search can help us express our perspective to other people. Early collaborations with creative communities have shown value in being able to create and share subjective experiences with others, resulting in feelings of being able to “break out of visually-similar echo chambers” or “see the world through another person’s eyes”. Even misalignment between model and human understanding of a concept frequently resulted in unexpected and inspiring connections for collaborators. Taken together, these findings point towards new ways of designing collaborative ML systems that embrace personal and collective subjectivity.

Conclusions and Future Work
Today, we’re open-sourcing the code to Mood Board Search, including three visual concepts made by our collaborators, and a Mood Board Search Python Library for people to tap the power of CAVs directly into their own websites and apps. While these tools are early-stage prototypes, we believe this capability can have a wide-range of applications from exploring unorganized image collections to externalizing ways of seeing into collaborative and shareable artifacts. Already, an experimental app by design invention studio Nord Projects, made using Mood Board Search, investigates the opportunities for running CAVs in camera, in real-time. In future work, we plan to use Mood Board Search to learn about new forms of human-machine collaboration and expand ML models and inputs — like text and audio — to allow even deeper subjective discoveries, regardless of medium.

If you’re interested in a demo of this work for your team or organization, email us at [email protected].

This blog presents research by (in alphabetical order): Kira Awadalla, Been Kim, Eva Kozanecka, Alison Lentz, Alice Moloney, Emily Reif, and Oliver Siy, in collaboration with design invention studio Nord Projects. We thank our co-author, Eva Kozanecka, our artist collaborators, Alexander Etchells, Tom Hatton, Rachel Maggart, the Imaging team at The British Library for their participation in beta previews, and Blaise Agüera y Arcas, Jess Holbrook, Fernanda Viegas, and Martin Wattenberg for their support of this research project.

Source: Google AI Blog

Vector-Quantized Image Modeling with Improved VQGAN

In recent years, natural language processing models have dramatically improved their ability to learn general-purpose representations, which has resulted in significant performance gains for a wide range of natural language generation and natural language understanding tasks. In large part, this has been accomplished through pre-training language models on extensive unlabeled text corpora.

This pre-training formulation does not make assumptions about input signal modality, which can be language, vision, or audio, among others. Several recent papers have exploited this formulation to dramatically improve image generation results through pre-quantizing images into discrete integer codes (represented as natural numbers), and modeling them autoregressively (i.e., predicting sequences one token at a time). In these approaches, a convolutional neural network (CNN) is trained to encode an image into discrete tokens, each corresponding to a small patch of the image. A second stage CNN or Transformer is then trained to model the distribution of encoded latent variables. The second stage can also be applied to autoregressively generate an image after the training. But while such models have achieved strong performance for image generation, few studies have evaluated the learned representation for downstream discriminative tasks (such as image classification).

In “Vector-Quantized Image Modeling with Improved VQGAN”, we propose a two-stage model that reconceives traditional image quantization techniques to yield improved performance on image generation and image understanding tasks. In the first stage, an image quantization model, called VQGAN, encodes an image into lower-dimensional discrete latent codes. Then a Transformer model is trained to model the quantized latent codes of an image. This approach, which we call Vector-quantized Image Modeling (VIM), can be used for both image generation and unsupervised image representation learning. We describe multiple improvements to the image quantizer and show that training a stronger image quantizer is a key component for improving both image generation and image understanding.

Vector-Quantized Image Modeling with ViT-VQGAN
One recent, commonly used model that quantizes images into integer tokens is the Vector-quantized Variational AutoEncoder (VQVAE), a CNN-based auto-encoder whose latent space is a matrix of discrete learnable variables, trained end-to-end. VQGAN is an improved version of this that introduces an adversarial loss to promote high quality reconstruction. VQGAN uses transformer-like elements in the form of non-local attention blocks, which allows it to capture distant interactions using fewer layers.

In our work, we propose taking this approach one step further by replacing both the CNN encoder and decoder with ViT. In addition, we introduce a linear projection from the output of the encoder to a low-dimensional latent variable space for lookup of the integer tokens. Specifically, we reduced the encoder output from a 768-dimension vector to a 32- or 8-dimension vector per code, which we found encourages the decoder to better utilize the token outputs, improving model capacity and efficiency.

Overview of the proposed ViT-VQGAN (left) and VIM (right), which, when working together, is capable of both image generation and image understanding. In the first stage, ViT-VQGAN converts images into discrete integers, which the autoregressive Transformer (Stage 2) then learns to model. Finally, the Stage 1 decoder is applied to these tokens to enable generation of high quality images from scratch.

With our trained ViT-VQGAN, images are encoded into discrete tokens represented by integers, each of which encompasses an 8x8 patch of the input image. Using these tokens, we train a decoder-only Transformer to predict a sequence of image tokens autoregressively. This two-stage model, VIM, is able to perform unconditioned image generation by simply sampling token-by-token from the output softmax distribution of the Transformer model.

VIM is also capable of performing class-conditioned generation, such as synthesizing a specific image of a given class (e.g., a dog or a cat). We extend the unconditional generation to class-conditioned generation by prepending a class-ID token before the image tokens during both training and sampling.

Uncurated set of dog samples from class-conditioned image generation trained on ImageNet. Conditioned classes: Irish terrier, Norfolk terrier, Norwich terrier, Yorkshire terrier, wire-haired fox terrier, Lakeland terrier.

To test the image understanding capabilities of VIM, we also fine-tune a linear projection layer to perform ImageNet classification, a standard benchmark for measuring image understanding abilities. Similar to ImageGPT, we take a layer output at a specific block, average over the sequence of token features (frozen) and insert a softmax layer (learnable) projecting averaged features to class logits. This allows us to capture intermediate features that provide more information useful for representation learning.

Experimental Results
We train all ViT-VQGAN models with a training batch size of 256 distributed across 128 CloudTPUv4 cores. All models are trained with an input image resolution of 256x256. On top of the pre-learned ViT-VQGAN image quantizer, we train Transformer models for unconditional and class-conditioned image synthesis and compare with previous work.

We measure the performance of our proposed methods for class-conditioned image synthesis and unsupervised representation learning on the widely used ImageNet benchmark. In the table below we demonstrate the class-conditioned image synthesis performance measured by the Fréchet Inception Distance (FID). Compared to prior work, VIM improves the FID to 3.07 (lower is better), a relative improvement of 58.6% over the VQGAN model (FID 7.35). VIM also improves the capacity for image understanding, as indicated by the Inception Score (IS), which goes from 188.6 to 227.4, a 20.6% improvement relative to VQGAN.

Model Acceptance

Validation data 1.0 1.62 235.0

DCTransformer 1.0 36.5 N/A
BigGAN 1.0 7.53 168.6
BigGAN-deep 1.0 6.84 203.6
IDDPM 1.0 12.3 N/A
ADM-G, 1.0 guid. 1.0 4.59 186.7
VQVAE-2 1.0 ~31 ~45

VQGAN 1.0 17.04 70.6
VQGAN 0.5 10.26 125.5
VQGAN 0.25 7.35 188.6
ViT-VQGAN (Ours) 1.0 4.17 175.1
ViT-VQGAN (Ours) 0.5 3.04 227.4
Fréchet Inception Distance (FID) comparison between different models for class-conditional image synthesis and Inception Score (IS) for image understanding, both on ImageNet with resolution 256x256. The acceptance rate shows results filtered by a ResNet-101 classification model, similar to the process in VQGAN.

After training a generative model, we test the learned image representations by fine-tuning a linear layer to perform ImageNet classification, a standard benchmark for measuring image understanding abilities. Our model outperforms previous generative models on the image understanding task, improving classification accuracy through linear probing (i.e., training a single linear classification layer, while keeping the rest of the model frozen) from 60.3% (iGPT-L) to 73.2%. These results showcase VIM’s strong generation results as well as image representation learning abilities.

We propose Vector-quantized Image Modeling (VIM), which pretrains a Transformer to predict image tokens autoregressively, where discrete image tokens are produced from improved ViT-VQGAN image quantizers. With our proposed improvements on image quantization, we demonstrate superior results on both image generation and understanding. We hope our results can inspire future work towards more unified approaches for image generation and understanding.

We would like to thank Xin Li, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, Yonghui Wu for the preparation of the VIM paper. We thank Wei Han, Yuan Cao, Jiquan Ngiam‎, Vijay Vasudevan, Zhifeng Chen and Claire Cui for helpful discussions and feedback, and others on the Google Research and Brain Team for support throughout this project.

Source: Google AI Blog

Extracting Skill-Centric State Abstractions from Value Functions

Advances in reinforcement learning (RL) for robotics have enabled robotic agents to perform increasingly complex tasks in challenging environments. Recent results show that robots can learn to fold clothes, dexterously manipulate a rubik’s cube, sort objects by color, navigate complex environments and walk on difficult, uneven terrain. But "short-horizon" tasks such as these, which require very little long-term planning and provide immediate failure feedback, are relatively easy to train compared to many tasks that may confront a robot in a real-world setting. Unfortunately, scaling such short-horizon skills to the abstract, long horizons of real-world tasks is difficult. For example, how would one train a robot capable of picking up objects to rearrange a room?

Hierarchical reinforcement learning (HRL), a popular way of solving this problem, has achieved some success in a variety of long-horizon RL tasks. HRL aims to solve such problems by reasoning over a bank of low-level skills, thus providing an abstraction for actions. However, the high-level planning problem can be further simplified by abstracting both states and actions. For example, consider a tabletop rearrangement task, where a robot is tasked with interacting with objects on a desk. Using recent advances in RL, imitation learning, and unsupervised skill discovery, it is possible to obtain a set of primitive manipulation skills such as opening or closing drawers, picking or placing objects, etc. However, even for the simple task of putting a block into the drawer, chaining these skills together is not straightforward. This may be attributed to a combination of (i) challenges with planning and reasoning over long horizons, and (ii) dealing with high dimensional observations while parsing the semantics and affordances of the scene, i.e., where and when the skill can be used.

In “Value Function Spaces: Skill-Centric State Abstractions for Long-Horizon Reasoning”, presented at ICLR 2022, we address the task of learning suitable state and action abstractions for long-range problems. We posit that a minimal, but complete, representation for a higher-level policy in HRL must depend on the capabilities of the skills available to it. We present a simple mechanism to obtain such a representation using skill value functions and show that such an approach improves long-horizon performance in both model-based and model-free RL and enables better zero-shot generalization.

Our method, VFS, can compose low-level primitives (left) to learn complex long-horizon behaviors (right).

Building a Value Function Space
The key insight motivating this work is that the abstract representation of actions and states is readily available from trained policies via their value functions. The notion of “value” in RL is intrinsically linked to affordances, in that the value of a state for skill reflects the probability of receiving a reward for successfully executing the skill. For any skill, its value function captures two key properties: 1) the preconditions and affordances of the scene, i.e., where and when the skill can be used, and 2) the outcome, which indicates whether the skill executed successfully when it was used.

Given a decision process with a finite set of k skills trained with sparse outcome rewards and their corresponding value functions, we construct an embedding space by stacking these skill value functions. This gives us an abstract representation that maps a state to a k-dimensional representation that we call the Value Function Space, or VFS for short. This representation captures functional information about the exhaustive set of interactions that the agent can have with the environment, and is thus a suitable state abstraction for downstream tasks.

Consider a toy example of the tabletop rearrangement setup discussed earlier, with the task of placing the blue object in the drawer. There are eight elementary actions in this environment. The bar plot on the right shows the values of each skill at any given time, and the graph at the bottom shows the evolution of these values over the course of the task.

Value functions corresponding to each skill (top-right; aggregated in bottom) capture functional information about the scene (top-left) and aid decision-making.

At the beginning, the values corresponding to the “Place on Counter” skill are high since the objects are already on the counter; likewise, the values corresponding to “Close Drawer” are high. Through the trajectory, when the robot picks up the blue cube, the corresponding skill value peaks. Similarly, the values corresponding to placing the objects in the drawer increase when the drawer is open and peak when the blue cube is placed inside it. All the functional information required to affect each transition and predict its outcome (success or failure) is captured by the VFS representation, and in principle, allows a high-level agent to reason over all the skills and chain them together — resulting in an effective representation of the observations.

Additionally, since VFS learns a skill-centric representation of the scene, it is robust to exogenous factors of variation, such as background distractors and appearances of task-irrelevant components of the scene. All configurations shown below are functionally equivalent — an open drawer with the blue cube in it, a red cube on the countertop, and an empty gripper — and can be interacted with identically, despite apparent differences.

The learned VFS representation can ignore task-irrelevant factors such as arm pose, distractor objects (green cube) and background appearance (brown desk).

Robotic Manipulation with VFS
This approach enables VFS to plan out complex robotic manipulation tasks. Take, for example, a simple model-based reinforcement learning (MBRL) algorithm that uses a simple one-step predictive model of the transition dynamics in value function space and randomly samples candidate skill sequences to select and execute the best one in a manner similar to the model-predictive control. Given a set of primitive pushing skills of the form “move Object A near Object B” and a high-level rearrangement task, we find that VFS can use MBRL to reliably find skill sequences that solve the high-level task.

A rollout of VFS performing a tabletop rearrangement task using a robotic arm. VFS can reason over a sequence of low-level primitives to achieve the desired goal configuration.

To better understand the attributes of the environment captured by VFS, we sample the VFS-encoded observations from a large number of independent trajectories in the robotic manipulation task and project them onto a two-dimensional axis using the t-SNE technique, which is useful for visualizing clusters in high-dimensional data. These t-SNE embeddings reveal interesting patterns identified and modeled by VFS. Looking at some of these clusters closely, we find that VFS can successfully capture information about the contents (objects) in the scene and affordances (e.g., a sponge can be manipulated when held by the robot’s gripper), while ignoring distractors like the relative positions of the objects on the table and the pose of the robotic arm. While these factors are certainly important to solve the task, the low-level primitives available to the robot abstract them away and hence, make them functionally irrelevant to the high-level controller.

Visualizing the 2D t-SNE projections of VFS embeddings show emergent clustering of equivalent configurations of the environment while ignoring task-irrelevant factors like arm pose.

Conclusions and Connections to Future Work
Value function spaces are representations built on value functions of underlying skills, enabling long-horizon reasoning and planning over skills. VFS is a compact representation that captures the affordances of the scene and task-relevant information while robustly ignoring distractors. Empirical experiments reveal that such a representation improves planning for model-based and model-free methods and enables zero-shot generalization. Going forward, this representation has the promise to continue improving along with the field of multitask reinforcement learning. The interpretability of VFS further enables integration into fields such as safe planning and grounding language models.

We thank our co-authors Sergey Levine, Ted Xiao, Alex Toshev, Peng Xu and Yao Lu for their contributions to the paper and feedback on this blog post. We also thank Tom Small for creating the informative visualizations used in this blog post.

Source: Google AI Blog

Pix2Seq: A New Language Interface for Object Detection

Object detection is a long-standing computer vision task that attempts to recognize and localize all objects of interest in an image. The complexity arises when trying to identify or localize all object instances while also avoiding duplication. Existing approaches, like Faster R-CNN and DETR, are carefully designed and highly customized in the choice of architecture and loss function. This specialization of existing systems has created two major barriers: (1) it adds complexity in tuning and training the different parts of the system (e.g., region proposal network, graph matching with GIOU loss, etc.), and (2), it can reduce the ability of a model to generalize, necessitating a redesign of the model for application to other tasks.

In “Pix2Seq: A Language Modeling Framework for Object Detection”, published at ICLR 2022, we present a simple and generic method that tackles object detection from a completely different perspective. Unlike existing approaches that are task-specific, we cast object detection as a language modeling task conditioned on the observed pixel inputs. We demonstrate that Pix2Seq achieves competitive results on the large-scale object detection COCO dataset compared to existing highly-specialized and well-optimized detection algorithms, and its performance can be further improved by pre-training the model on a larger object detection dataset. To encourage further research in this direction, we are also excited to release to the broader research community Pix2Seq’s code and pre-trained models along with an interactive demo.

Pix2Seq Overview
Our approach is based on the intuition that if a neural network knows where and what the objects in an image are, one could simply teach it how to read them out. By learning to “describe” objects, the model can learn to ground the descriptions on pixel observations, leading to useful object representations. Given an image, the Pix2Seq model outputs a sequence of object descriptions, where each object is described using five discrete tokens: the coordinates of the bounding box’s corners [ymin, xmin, ymax, xmax] and a class label.

Pix2Seq framework for object detection. The neural network perceives an image, and generates a sequence of tokens for each object, which correspond to bounding boxes and class labels.

With Pix2Seq, we propose a quantization and serialization scheme that converts bounding boxes and class labels into sequences of discrete tokens (similar to captions), and leverage an encoder-decoder architecture to perceive pixel inputs and generate the sequence of object descriptions. The training objective function is simply the maximum likelihood of tokens conditioned on pixel inputs and the preceding tokens.

Sequence Construction from Object Descriptions
In commonly used object detection datasets, images have variable numbers of objects, represented as sets of bounding boxes and class labels. In Pix2Seq, a single object, defined by a bounding box and class label, is represented as [ymin, xmin, ymax, xmax, class]. However, typical language models are designed to process discrete tokens (or integers) and are unable to comprehend continuous numbers. So, instead of representing image coordinates as continuous numbers, we normalize the coordinates between 0 and 1 and quantize them into one of a few hundred or thousand discrete bins. The coordinates are then converted into discrete tokens as are the object descriptions, similar to image captions, which in turn can then be interpreted by the language model. The quantization process is achieved by multiplying the normalized coordinate (e.g., ymin) by the number of bins minus one, and rounding it to the nearest integer (the detailed process can be found in our paper).

Quantization of the coordinates of the bounding boxes with different numbers of bins on a 480 × 640 image. With a small number of bins/tokens, such as 500 bins (∼1 pixel/bin), it achieves high precision even for small objects.

After quantization, the object annotations provided with each training image are ordered into a sequence of discrete tokens (shown below). Since the order of the objects does not matter for the detection task per se, we randomize the order of objects each time an image is shown during training. We also append an End of Sequence (EOS) token at the end as​​ different images often have different numbers of objects, and hence sequence lengths.

The bounding boxes and class labels for objects detected in the image on the left are represented in the sequences shown on the right. A random object ordering strategy is used in our work but other approaches to ordering could also be used.

The Model Architecture, Objective Function, and Inference
We treat the sequences that we constructed from object descriptions as a “dialect” and address the problem via a powerful and general language model with an image encoder and autoregressive language encoder. Similar to language modeling, Pix2Seq is trained to predict tokens, given an image and preceding tokens, with a maximum likelihood loss. At inference time, we sample tokens from model likelihood. The sampled sequence ends when the EOS token is generated. Once the sequence is generated, we split it into chunks of 5 tokens for extracting and de-quantizing the object descriptions (i.e., obtaining the predicted bounding boxes and class labels). It is worth noting that both the architecture and loss function are task-agnostic in that they don’t assume prior knowledge about object detection (e.g., bounding boxes). We describe how we can incorporate task-specific prior knowledge with a sequence augmentation technique in our paper.

Despite its simplicity, Pix2Seq achieves impressive empirical performance on benchmark datasets. Specifically, we compare our method with well established baselines, Faster R-CNN and DETR, on the widely used COCO dataset and demonstrate that it achieves competitive average precision (AP) results.

Pix2Seq achieves competitive AP results compared to existing systems that require specialization during model design, while being significantly simpler. The best performing Pix2Seq model achieved an AP score of 45.

Since our approach incorporates minimal inductive bias or prior knowledge of the object detection task into the model design, we further explore how pre-training the model using the large-scale object detection COCO dataset can impact its performance. Our results indicate that this training strategy (along with using bigger models) can further boost performance.

The average precision of the Pix2Seq model with pre-training followed by fine-tuning. The best performing Pix2Seq model without pre-training achieved an AP score of 45. When the model is pre-trained, we see an 11% improvement with an AP score of 50.

Pix2Seq can detect objects in densely populated and complex scenes, such as those shown below.

Example complex and densely populated scenes labeled by a trained Pix2Seq model. Try it out here.

Conclusion and Future Work
With Pix2Seq, we cast object detection as a language modeling task conditioned on pixel inputs for which the model architecture and loss function are generic, and have not been engineered specifically for the detection task. One can, therefore, readily extend this framework to different domains or applications, where the output of the system can be represented by a relatively concise sequence of discrete tokens (e.g., keypoint detection, image captioning, visual question answering), or incorporate it into a perceptual system supporting general intelligence, for which it provides a language interface to a wide range of vision and language tasks. We also hope that the release of our Pix2Seq’s code, pre-trained models and interactive demo will inspire further research in this direction.

This post reflects the combined work with our co-authors: Saurabh Saxena, Lala Li, Geoffrey Hinton. We would also like to thank Tom Small for the visualization of the Pix2Seq illustration figure.

Source: Google AI Blog

Offline Optimization for Architecting Hardware Accelerators

Advances in machine learning (ML) often come with advances in hardware and computing systems. For example, the growth of ML-based approaches in solving various problems in vision and language has led to the development of application-specific hardware accelerators (e.g., Google TPUs and Edge TPUs). While promising, standard procedures for designing accelerators customized towards a target application require manual effort to devise a reasonably accurate simulator of hardware, followed by performing many time-intensive simulations to optimize the desired objective (e.g., optimizing for low power usage or latency when running a particular application). This involves identifying the right balance between total amount of compute and memory resources and communication bandwidth under various design constraints, such as the requirement to meet an upper bound on chip area usage and peak power. However, designing accelerators that meet these design constraints is often result in infeasible designs. To address these challenges, we ask: “Is it possible to train an expressive deep neural network model on large amounts of existing accelerator data and then use the learned model to architect future generations of specialized accelerators, eliminating the need for computationally expensive hardware simulations?

In “Data-Driven Offline Optimization for Architecting Hardware Accelerators”, accepted at ICLR 2022, we introduce PRIME, an approach focused on architecting accelerators based on data-driven optimization that only utilizes existing logged data (e.g., data leftover from traditional accelerator design efforts), consisting of accelerator designs and their corresponding performance metrics (e.g., latency, power, etc) to architect hardware accelerators without any further hardware simulation. This alleviates the need to run time-consuming simulations and enables reuse of data from past experiments, even when the set of target applications changes (e.g., an ML model for vision, language, or other objective), and even for unseen but related applications to the training set, in a zero-shot fashion. PRIME can be trained on data from prior simulations, a database of actually fabricated accelerators, and also a database of infeasible or failed accelerator designs1. This approach for architecting accelerators — tailored towards both single- and multi-applications — improves performance upon state-of-the-art simulation-driven methods by about 1.2x-1.5x, while considerably reducing the required total simulation time by 93% and 99%, respectively. PRIME also architects effective accelerators for unseen applications in a zero-shot setting, outperforming simulation-based methods by 1.26x.

PRIME uses logged accelerator data, consisting of both feasible and infeasible accelerators, to train a conservative model, which is used to design accelerators while meeting design constraints. PRIME architects accelerators with up to 1.5x smaller latency, while reducing the required hardware simulation time by up to 99%.

The PRIME Approach for Architecting Accelerators
Perhaps the simplest possible way to use a database of previously designed accelerators for hardware design is to use supervised machine learning to train a prediction model that can predict the performance objective for a given accelerator as input. Then, one could potentially design new accelerators by optimizing the performance output of this learned model with respect to the input accelerator design. Such an approach is known as model-based optimization. However, this simple approach has a key limitation: it assumes that the prediction model can accurately predict the cost for every accelerator that we might encounter during optimization! It is well established that most prediction models trained via supervised learning misclassify adversarial examples that “fool” the learned model into predicting incorrect values. Similarly, it has been shown that even optimizing the output of a supervised model finds adversarial examples that look promising under the learned model2, but perform terribly under the ground truth objective.

To address this limitation, PRIME learns a robust prediction model that is not prone to being fooled by adversarial examples (that we will describe shortly), which would be otherwise found during optimization. One can then simply optimize this model using any standard optimizer to architect simulators. More importantly, unlike prior methods, PRIME can also utilize existing databases of infeasible accelerators to learn what not to design. This is done by augmenting the supervised training of the learned model with additional loss terms that specifically penalize the value of the learned model on the infeasible accelerator designs and adversarial examples during training. This approach resembles a form of adversarial training.

In principle, one of the central benefits of a data-driven approach is that it should enable learning highly expressive and generalist models of the optimization objective that generalize over target applications, while also potentially being effective for new unseen applications for which a designer has never attempted to optimize accelerators. To train PRIME so that it generalizes to unseen applications, we modify the learned model to be conditioned on a context vector that identifies a given neural net application we wish to accelerate (as we discuss in our experiments below, we choose to use high-level features of the target application: such as number of feed-forward layers, number of convolutional layers, total parameters, etc. to serve as the context), and train a single, large model on accelerator data for all applications designers have seen so far. As we will discuss below in our results, this contextual modification of PRIME enables it to optimize accelerators both for multiple, simultaneous applications and new unseen applications in a zero-shot fashion.

Does PRIME Outperform Custom-Engineered Accelerators?
We evaluate PRIME on a variety of actual accelerator design tasks. We start by comparing the optimized accelerator design architected by PRIME targeted towards nine applications to the manually optimized EdgeTPU design. EdgeTPU accelerators are primarily optimized towards running applications in image classification, particularly MobileNetV2, MobileNetV3 and MobileNetEdge. Our goal is to check if PRIME can design an accelerator that attains a lower latency than a baseline EdgeTPU accelerator3, while also constraining the chip area to be under 27 mm2 (the default for the EdgeTPU accelerator). Shown below, we find that PRIME improves latency over EdgeTPU by 2.69x (up to 11.84x in t-RNN Enc), while also reducing the chip area usage by 1.50x (up to 2.28x in MobileNetV3), even though it was never trained to reduce chip area! Even on the MobileNet image-classification models, for which the custom-engineered EdgeTPU accelerator was optimized, PRIME improves latency by 1.85x.

Comparing latencies (lower is better) of accelerator designs suggested by PRIME and EdgeTPU for single-model specialization.
The chip area (lower is better) reduction compared to a baseline EdgeTPU design for single-model specialization.

Designing Accelerators for New and Multiple Applications, Zero-Shot
We now study how PRIME can use logged accelerator data to design accelerators for (1) multiple applications, where we optimize PRIME to design a single accelerator that works well across multiple applications simultaneously, and in a (2) zero-shot setting, where PRIME must generate an accelerator for new unseen application(s) without training on any data from such applications. In both settings, we train the contextual version of PRIME, conditioned on context vectors identifying the target applications and then optimize the learned model to obtain the final accelerator. We find that PRIME outperforms the best simulator-driven approach in both settings, even when very limited data is provided for training for a given application but many applications are available. Specifically in the zero-shot setting, PRIME outperforms the best simulator-driven method we compared to, attaining a reduction of 1.26x in latency. Further, the difference in performance increases as the number of training applications increases.

The average latency (lower is better) of test applications under zero-shot setting compared to a state-of-the-art simulator-driven approach. The text on top of each bar shows the set of training applications.

Closely Analyzing an Accelerator Designed by PRIME
To provide more insight to hardware architecture, we examine the best accelerator designed by PRIME and compare it to the best accelerator found by the simulator-driven approach. We consider the setting where we need to jointly optimize the accelerator for all nine applications, MobileNetEdge, MobileNetV2, MobileNetV3, M4, M5, M64, t-RNN Dec, and t-RNN Enc, and U-Net, under a chip area constraint of 100 mm2. We find that PRIME improves latency by 1.35x over the simulator-driven approach.

Per application latency (lower is better) for the best accelerator design suggested by PRIME and state-of-the-art simulator-driven approach for a multi-task accelerator design. PRIME reduces the average latency across all nine applications by 1.35x over the simulator-driven method.

As shown above, while the latency of the accelerator designed by PRIME for MobileNetEdge, MobileNetV2, MobileNetV3, M4, t-RNN Dec, and t-RNN Enc are better, the accelerator found by the simulation-driven approach yields a lower latency in M5, M6, and U-Net. By closely inspecting the accelerator configurations, we find that PRIME trades compute (64 cores for PRIME vs. 128 cores for the simulator-driven approach) for larger Processing Element (PE) memory size (2,097,152 bytes vs. 1,048,576 bytes). These results show that PRIME favors PE memory size to accommodate the larger memory requirements in t-RNN Dec and t-RNN Enc, where large reductions in latency were possible. Under a fixed area budget, favoring larger on-chip memory comes at the expense of lower compute power in the accelerator. This reduction in the accelerator's compute power leads to higher latency for the models with large numbers of compute operations, namely M5, M6, and U-Net.

The efficacy of PRIME highlights the potential for utilizing the logged offline data in an accelerator design pipeline. A likely avenue for future work is to scale this approach across an array of applications, where we expect to see larger gains because simulator-driven approaches would need to solve a complex optimization problem, akin to searching for needle in a haystack, whereas PRIME can benefit from generalization of the surrogate model. On the other hand, we would also note that PRIME outperforms prior simulator-driven methods we utilize and this makes it a promising candidate to be used within a simulator-driven method. More generally, training a strong offline optimization algorithm on offline datasets of low-performing designs can be a highly effective ingredient in at the very least, kickstarting hardware design, versus throwing out prior data. Finally, given the generality of PRIME, we hope to use it for hardware-software co-design, which exhibits a large search space but plenty of opportunity for generalization. We have also released both the code for training PRIME and the dataset of accelerators.

We thank our co-authors Sergey Levine, Kevin Swersky, and Milad Hashemi for their advice, thoughts and suggestions. We thank James Laudon, Cliff Young, Ravi Narayanaswami, Berkin Akin, Sheng-Chun Kao, Samira Khan, Suvinay Subramanian, Stella Aslibekyan, Christof Angermueller, and Olga Wichrowskafor for their help and support, and Sergey Levine for feedback on this blog post. In addition, we would like to extend our gratitude to the members of “Learn to Design Accelerators”, “EdgeTPU”, and the Vizier team for providing invaluable feedback and suggestions. We would also like to thank Tom Small for the animated figure used in this post.

1The infeasible accelerator designs stem from build errors in silicon or compilation/mapping failures. 
2This is akin to adversarial examples in supervised learning – these examples are close to the data points observed in the training dataset, but are misclassified by the classifier. 
3The performance metrics for the baseline EdgeTPU accelerator are extracted from an industry-based hardware simulator tuned to match the performance of the actual hardware. 
4These are proprietary object-detection models, and we refer to them as M4 (indicating Model 4), M5, and M6 in the paper. 

Source: Google AI Blog

Do Wide and Deep Networks Learn the Same Things?

A common practice to improve a neural network’s performance and tailor it to available computational resources is to adjust the architecture depth and width. Indeed, popular families of neural networks, including EfficientNet, ResNet and Transformers, consist of a set of architectures of flexible depths and widths. However, beyond the effect on accuracy, there is limited understanding of how these fundamental choices of architecture design affect the model, such as the impact on its internal representations.

In “Do Wide and Deep Networks Learn the Same Things? Uncovering How Neural Network Representations Vary with Width and Depth”, we perform a systematic study of the similarity between wide and deep networks from the same architectural family through the lens of their hidden representations and final outputs. In very wide or very deep models, we find a characteristic block structure in their internal representations, and establish a connection between this phenomenon and model overparameterization. Comparisons across models demonstrate that those without the block structure show significant similarity between representations in corresponding layers, but those containing the block structure exhibit highly dissimilar representations. These properties of the internal representations in turn translate to systematically different errors at the class and example levels for wide and deep models when they are evaluated on the same test set.

Comparing Representation Similarity with CKA
We extended prior work on analyzing representations by leveraging our previously developed Centered Kernel Alignment (CKA) technique, which provides a robust, scalable way to determine the similarity between the representations learned by any pair of neural network layers. CKA takes as input the representations (i.e., the activation matrices) from two layers, and outputs a similarity score between 0 (not at all similar) and 1 (identical representations).

We apply CKA to a family of ResNets of varying depths and widths, trained on common benchmark datasets (CIFAR-10, CIFAR-100 and ImageNet), and use representation heatmaps to illustrate the results. The x and y axes of each heatmap index the layers of the model(s) in consideration, going from input to output, and each entry (i, j) is the CKA similarity score between layer i and layer j.

We use CKA to compute the representation similarity for all pairs of layers within a single model (i.e., when network 1 and network 2 are identical), and across models (i.e., when network 1 and network 2 are trained with different random initializations, or have different architectures altogether).

Below is an example of the resulting heatmap when we compare representations of each layer to every other layer within a single ResNet of depth 26 and width multiplier 1. In the design convention used here, the stated depth only refers to the number of convolutional layers in the network, but we analyze all layers present, and the width multiplier applies to the number of filters in each convolution. Notice the checkerboard pattern in the heatmap, which is caused by skip connections (shortcuts between layers) in the architecture.

The Emergence of the Block Structure
What stands out from the representation heatmaps of deeper or wider networks is the emergence of a large set of consecutive layers with highly similar representations, which appears in the heatmaps as a yellow square (i.e., a region with high CKA scores). This phenomenon, which we call the block structure, suggests that the underlying layers may not be as efficient at progressively refining the network’s representations as we expect. Indeed, we show that the task performance becomes stagnant inside the block structure, and that it is possible to prune some underlying layers without affecting the final performance.

Block structure — a large, contiguous set of layers with highly similar representations — emerges with increasing width or depth. Each heatmap panel shows the CKA similarity between all pairs of layers within a single neural network. While its size and position can vary across different training runs, the block structure is a robust phenomenon that arises consistently in larger models.

With additional experiments, we show that the block structure has less to do with the absolute model size, than with the size of the model relative to the size of the training dataset. As we reduce the training dataset size, the block structure starts to appear in shallower and narrower networks:

With increasing network width (towards the right along each row) and decreasing dataset size (down each column), the relative model capacity (with respect to a given task) is effectively inflated, and the block structure begins to appear in smaller models.

Through further analysis, we are also able to demonstrate that the block structure arises from preserving and propagating the dominant principal components of its underlying representations. Refer to our paper for more details.

Comparing Representations Across Models
Going further, we study the implications of depth and width on representations across models of different random initializations and different architectures, and find that the presence of block structure makes a significant difference in this context as well. Despite having different architectures, wide and deep models without the block structure do exhibit representation similarity with each other, with corresponding layers broadly being of the same proportional depth in the model. However, when the block structure is present, its representations are unique to each model. This suggests that despite having similar overall performance, each wide or deep model with the block structure picks up a unique mapping from the input to the output.

For smaller models (e.g., ResNet-38 1×), CKA across different initializations (off the diagonal) closely resembles CKA within a single model (on the diagonal). In contrast, representations within the block structure of wider and deeper models (e.g., ResNet-38 10×, ResNet-164 1×) are highly dissimilar across training runs.

Error Analysis of Wide and Deep Models
Having explored the properties of the learned representations of wide and deep models, we next turn to understanding how they influence the diversity of the output predictions. We train populations of networks of different architectures and determine on which test set examples each architecture configuration tends to make errors.

On both CIFAR-10 and ImageNet datasets, wide and deep models that have the same average accuracy still demonstrate statistically significant differences in example-level predictions. The same observation holds for class-level errors on ImageNet, with wide models exhibiting a small advantage in identifying classes corresponding to scenes, and deep networks being relatively more accurate on consumer goods.

Per-class differences on ImageNet between models with increased width (y-axis) or depth (x-axis). Orange dots reflect differences between two sets of 50 different random initializations of ResNet-83 (1×).

In studying the effects of depth and width on internal representations, we uncover a block structure phenomenon, and demonstrate its connection to model capacity. We also show that wide and deep models exhibit systematic output differences at class and example levels. Check out the paper for full details on these results and additional insights! We’re excited about the many interesting open questions these findings suggest, such as how the block structure arises during training, whether the phenomenon occurs in domains beyond image classification, and ways these insights on internal representations can inform model efficiency and generalization.

This is a joint work with Maithra Raghu and Simon Kornblith. We would like to thank Tom Small for the visualizations of the representation heatmap.

Source: Google AI Blog