Google at CVPR 2022

This week marks the beginning of the premier annual Computer Vision and Pattern Recognition conference (CVPR 2022), held both in-person in New Orleans, LA and virtually. As a leader in computer vision research and a Platinum Sponsor, Google will have a strong presence across CVPR 2022 with over 80 papers being presented at the main conference and active involvement in a number of conference workshops and tutorials.

If you are attending CVPR this year, please stop by our booth and chat with our researchers who are actively exploring the latest machine learning techniques for application to various areas of machine perception. Our researchers will also be available to talk about and demo several recent efforts, including on-device ML applications with MediaPipe, the Auto Arborist Dataset for urban forest monitoring, and much more.

You can also learn more about our research being presented at CVPR 2022 in the list below (Google affiliations in bold).

Organizing Committee

Tutorials Chairs
Include: Boqing Gong

Website Chairs
Include: AJ Piergiovanni

Area Chairs
Include: Alireza Fathi, Cordelia Schmid, Deqing Sun, Jonathan Barron, Michael Ryoo, Supasorn Suwajanakorn, Susanna Ricco

Diversity, Equity, and Inclusion Chairs
Include: Noah Snavely

Panel Discussion: Embodied Computer Vision
Panelists include: Michael Ryoo

Publications

Learning to Prompt for Continual Learning (see blog post)
Zifeng Wang*, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, Tomas Pfister

GCR: Gradient Coreset Based Replay Buffer Selection for Continual Learning
Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, Pradeep Shenoy

Zero-Shot Text-Guided Object Generation with Dream Fields
Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, Ben Poole

Towards End-to-End Unified Scene Text Detection and Layout Analysis
Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, Michalis Raptis

FLOAT: Factorized Learning of Object Attributes for Improved Multi-object Multi-part Scene Parsing
Rishubh Singh, Pranav Gupta, Pradeep Shenoy, Ravikiran Sarvadevabhatla

LOLNerf: Learn from One Look
Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagun, Andrea Tagliasacchi

Photorealistic Monocular 3D Reconstruction of Humans Wearing Clothing
Thiemo Alldieck, Mihai Zanfir, Cristian Sminchisescu

Learning Local Displacements for Point Cloud Completion
Yida Wang, David Joseph Tan, Nassir Navab, Federico Tombari

Density-Preserving Deep Point Cloud Compression
Yun He, Xinlin Ren, Danhang Tang, Yinda Zhang, Xiangyang Xue, Yanwei Fu

CMT-DeepLab: Clustering Mask Transformers for Panoptic Segmentation
Qihang Yu*, Huiyu Wang, Dahun Kim, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, Liang-Chieh Chen

Deformable Sprites for Unsupervised Video Decomposition
Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, Noah Snavely

Learning with Neighbor Consistency for Noisy Labels
Ahmet Iscen, Jack Valmadre, Anurag Arnab, Cordelia Schmid

Multiview Transformers for Video Recognition
Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, Cordelia Schmid

Kubric: A Scalable Dataset Generator
Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan*, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, Andrea Tagliasacchi

3D Moments from Near-Duplicate Photos
Qianqian Wang, Zhengqi Li, David Salesin, Noah Snavely, Brian Curless, Janne Kontkanen

Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields
Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, Peter Hedman

RegNeRF: Regularizing Neural Radiance Fields for View Synthesis from Sparse Inputs
Michael Niemeyer*, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, Noha Radwan*

Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields
Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, Pratul P. Srinivasan

IRON: Inverse Rendering by Optimizing Neural SDFs and Materials from Photometric Images
Kai Zhang, Fujun Luan, Zhengqi Li, Noah Snavely

MAXIM: Multi-Axis MLP for Image Processing
Zhengzhong Tu*, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, Yinxiao Li

Restormer: Efficient Transformer for High-Resolution Image Restoration
Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang

Burst Image Restoration and Enhancement
Akshay Dudhane, Syed Waqas Zamir, Salman Khan, Fahad Shahbaz Khan, Ming-Hsuan Yang

Neural RGB-D Surface Reconstruction
Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, Justus Thies

Scene Representation Transformer: Geometry-Free Novel View Synthesis Through Set-Latent Scene Representations
Mehdi S. M. Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan*, Suhani Vora, Mario Lučić, Daniel Duckworth, Alexey Dosovitskiy*, Jakob Uszkoreit*, Thomas Funkhouser, Andrea Tagliasacchi*

ZebraPose: Coarse to Fine Surface Encoding for 6DoF Object Pose Estimation
Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, Federico Tombari

MetaPose: Fast 3D Pose from Multiple Views without 3D Supervision
Ben Usman, Andrea Tagliasacchi, Kate Saenko, Avneesh Sud

GPV-Pose: Category-Level Object Pose Estimation via Geometry-Guided Point-wise Voting
Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xiangyang Ji, Nassir Navab, Federico Tombari

Rethinking Deep Face Restoration
Yang Zhao*, Yu-Chuan Su, Chun-Te Chu, Yandong Li, Marius Renn, Yukun Zhu, Changyou Chen, Xuhui Jia

Transferability Metrics for Selecting Source Model Ensembles
Andrea Agostinelli, Jasper Uijlings, Thomas Mensink, Vittorio Ferrari

Robust Fine-Tuning of Zero-Shot Models
Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, Ludwig Schmidt

Block-NeRF: Scalable Large Scene Neural View Synthesis
Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, Henrik Kretzschmar

Light Field Neural Rendering
Mohammad Suhail*, Carlos Esteves, Leonid Sigal, Ameesh Makadia

Transferability Estimation Using Bhattacharyya Class Separability
Michal Pándy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, Thomas Mensink

Matching Feature Sets for Few-Shot Image Classification
Arman Afrasiyabi, Hugo Larochelle, Jean-François Lalonde, Christian Gagné

Which Model to Transfer? Finding the Needle in the Growing Haystack
Cedric Renggli, André Susano Pinto, Luka Rimanic, Joan Puigcerver, Carlos Riquelme, Ce Zhang, Mario Lučić

Auditing Privacy Defenses in Federated Learning via Generative Gradient Leakage
Zhuohang Li, Jiaxin Zhang, Luyang Liu, Jian Liu

Estimating Example Difficulty Using Variance of Gradients
Chirag Agarwal, Daniel D’souza, Sara Hooker

More Than Words: In-the-Wild Visually-Driven Prosody for Text-to-Speech (see blog post)
Michael Hassid, Michelle Tadmor Ramanovich, Brendan Shillingford, Miaosen Wang, Ye Jia, Tal Remez

Robust Outlier Detection by De-Biasing VAE Likelihoods
Kushal Chauhan, Barath Mohan U, Pradeep Shenoy, Manish Gupta, Devarajan Sridharan

Deep 3D-to-2D Watermarking: Embedding Messages in 3D Meshes and Extracting Them from 2D Renderings
Innfarn Yoo, Huiwen Chang, Xiyang Luo, Ondrej Stava, Ce Liu*, Peyman Milanfar, Feng Yang

Knowledge Distillation: A Good Teacher Is Patient and Consistent
Lucas Beyer, Xiaohua Zhai, Amélie Royer*, Larisa Markeeva*, Rohan Anil, Alexander Kolesnikov

Urban Radiance Fields
Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas Funkhouser, Vittorio Ferrari

Manifold Learning Benefits GANs
Yao Ni, Piotr Koniusz, Richard Hartley, Richard Nock

MaskGIT: Masked Generative Image Transformer
Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu*, William T. Freeman

InOut: Diverse Image Outpainting via GAN Inversion
Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, Ming-Hsuan Yang

Scaling Vision Transformers (see blog post)
Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, Lucas Beyer

Fine-Tuning Image Transformers Using Learnable Memory
Mark Sandler, Andrey Zhmoginov, Max Vladymyrov, Andrew Jackson

PokeBNN: A Binary Pursuit of Lightweight Accuracy
Yichi Zhang*, Zhiru Zhang, Lukasz Lew

Bending Graphs: Hierarchical Shape Matching Using Gated Optimal Transport
Mahdi Saleh, Shun-Cheng Wu, Luca Cosmo, Nassir Navab, Benjamin Busam, Federico Tombari

Uncertainty-Aware Deep Multi-View Photometric Stereo
Berk Kaya, Suryansh Kumar, Carlos Oliveira, Vittorio Ferrari, Luc Van Gool

Depth-Supervised NeRF: Fewer Views and Faster Training for Free
Kangle Deng, Andrew Liu, Jun-Yan Zhu, Deva Ramanan

Dense Depth Priors for Neural Radiance Fields from Sparse Input Views
Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, Matthias Nießner

Trajectory Optimization for Physics-Based Reconstruction of 3D Human Pose from Monocular Video
Erik Gärtner, Mykhaylo Andriluka, Hongyi Xu, Cristian Sminchisescu

Differentiable Dynamics for Articulated 3D Human Motion Reconstruction
Erik Gärtner, Mykhaylo Andriluka, Erwin Coumans, Cristian Sminchisescu

Panoptic Neural Fields: A Semantic Object-Aware Neural Scene Representation
Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J. Guibas, Andrea Tagliasacchi, Frank Dellaert, Thomas Funkhouser

Pyramid Adversarial Training Improves ViT Performance
Charles Herrmann, Kyle Sargent, Lu Jiang, Ramin Zabih, Huiwen Chang, Ce Liu*, Dilip Krishnan, Deqing Sun

Proper Reuse of Image Classification Features Improves Object Detection
Cristina Vasconcelos, Vighnesh Birodkar, Vincent Dumoulin

SOMSI: Spherical Novel View Synthesis with Soft Occlusion Multi-Sphere Images
Tewodros Habtegebrial, Christiano Gava, Marcel Rogge, Didier Stricker, Varun Jampani

TubeFormer-DeepLab: Video Mask Transformer
Dahun Kim, Jun Xie, Huiyu Wang, Siyuan Qiao, Qihang Yu, Hong-Seok Kim, Hartwig Adam, In So Kweon, Liang-Chieh Chen

Contextualized Spatio-Temporal Contrastive Learning with Self-Supervision
Liangzhe Yuan, Rui Qian*, Yin Cui, Boqing Gong, Florian Schroff, Ming-Hsuan Yang, Hartwig Adam, Ting Liu

When Does Contrastive Visual Representation Learning Work?
Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, Serge Belongie

Less Is More: Generating Grounded Navigation Instructions from Landmarks
Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, Peter Anderson

Forecasting Characteristic 3D Poses of Human Actions
Christian Diller, Thomas Funkhouser, Angela Dai

BEHAVE: Dataset and Method for Tracking Human Object Interactions
Bharat Lal Bhatnagar, Xianghui Xie, Ilya A. Petrov, Cristian Sminchisescu, Christian Theobalt, Gerard Pons-Moll

Motion-from-Blur: 3D Shape and Motion Estimation of Motion-Blurred Objects in Videos
Denys Rozumnyi, Martin R. Oswald, Vittorio Ferrari, Marc Pollefeys

End-to-End Generative Pretraining for Multimodal Video Captioning (see blog post)
Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, Cordelia Schmid

Uncertainty-Aware Adaptation for Self-Supervised 3D Human Pose Estimation
Jogendra Nath Kundu, Siddharth Seth, Pradyumna YM, Varun Jampani, Anirban Chakraborty, R. Venkatesh Babu

Learning ABCs: Approximate Bijective Correspondence for Isolating Factors of Variation with Weak Supervision
Kieran A. Murphy, Varun Jampani, Srikumar Ramalingam, Ameesh Makadia

HumanNeRF: Free-Viewpoint Rendering of Moving People from Monocular Video
Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, Ira Kemelmacher-Shlizerman

Deblurring via Stochastic Refinement
Jay Whang*, Mauricio Delbracio, Hossein Talebi, Chitwan Saharia, Alexandros G. Dimakis, Peyman Milanfar

NeRF in the Dark: High Dynamic Range View Synthesis from Noisy Raw Images
Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron

CoNeRF: Controllable Neural Radiance Fields
Kacper Kania, Kwang Moo Yi, Marek Kowalski, Tomasz Trzciński, Andrea Tagliasacchi

A Conservative Approach for Unbiased Learning on Unknown Biases
Myeongho Jeon, Daekyung Kim, Woochul Lee, Myungjoo Kang, Joonseok Lee

DeepFusion: Lidar-Camera Deep Fusion for Multi-Modal 3D Object Detection (see blog post)
Yingwei Li*, Adams Wei Yu, Tianjian Meng, Ben Caine, Jiquan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V. Le, Alan Yuille, Mingxing Tan

Video Frame Interpolation Transformer
Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, Ming-Hsuan Yang

Global Matching with Overlapping Attention for Optical Flow Estimation
Shiyu Zhao, Long Zhao, Zhixing Zhang, Enyu Zhou, Dimitris Metaxas

LiT: Zero-Shot Transfer with Locked-image Text Tuning (see blog post)
Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, Lucas Beyer

Are Multimodal Transformers Robust to Missing Modality?
Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, Xi Peng

3D-VField: Adversarial Augmentation of Point Clouds for Domain Generalization in 3D Object Detection
Alexander Lehner, Stefano Gasperini, Alvaro Marcos-Ramiro, Michael Schmidt, Mohammad-Ali Nikouei Mahani, Nassir Navab, Benjamin Busam, Federico Tombari

SHIFT: A Synthetic Driving Dataset for Continuous Multi-Task Domain Adaptation
Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, Fisher Yu

H4D: Human 4D Modeling by Learning Neural Compositional Representation
Boyan Jiang, Yinda Zhang, Xingkui Wei, Xiangyang Xue, Yanwei Fu

Gravitationally Lensed Black Hole Emission Tomography
Aviad Levis, Pratul P. Srinivasan, Andrew A. Chael, Ren Ng, Katherine L. Bouman

Deep Saliency Prior for Reducing Visual Distraction
Kfir Aberman, Junfeng He, Yossi Gandelsman, Inbar Mosseri, David E. Jacobs, Kai Kohlhoff, Yael Pritch, Michael Rubinstein

The Auto Arborist Dataset: A Large-Scale Benchmark for Multiview Urban Forest Monitoring Under Domain Shift
Sara Beery, Guanhang Wu, Trevor Edwards, Filip Pavetic, Bo Majewski, Shreyasee Mukherjee, Stanley Chan, John Morgan, Vivek Rathod, Jonathan Huang

Workshops

Ethical Considerations in Creative Applications of Computer Vision
Chairs and Advisors: Negar Rostamzadeh, Fernando Diaz, Emily Denton, Mark Diaz, Jason Baldridge

Dynamic Neural Networks Meet Computer Vision Organizers
Invited Speaker: Barret Zoph

Precognition: Seeing Through the Future
Organizer: Utsav Prabhu
Invited Speaker: Sella Nevo

Computer Vision in the Built Environment for the Design, Construction, and Operation of Buildings
Invited Speakers: Thomas Funkhouser, Federico Tombari

Neural Architecture Search: Lightweight NAS Challenge
Invited Speaker: Barret Zoph

Transformers in Vision
Organizer: Lucas Beyer
Invited Speakers and Panelists: Alexander Kolesnikov, Mathilde Caron, Arsha Nagrani, Lucas Beyer

Challenge on Learned Image Compression
Organizers: George Toderici, Johannes Balle, Eirikur Agustsson, Nick Johnston, Fabian Mentzer, Luca Versari
Invited Speaker: Debargha Mukherjee

Embodied AI
Organizers: Anthony Francis, Sören Pirk, Alex Ku, Fei Xia, Peter Anderson
Scientific Advisory Board Members: Alexander Toshev, Jie Tan
Invited Speaker: Carolina Parada

Sight and Sound
Organizers: Arsha Nagrani, William Freeman

New Trends in Image Restoration and Enhancement
Organizers: Ming-Hsuan Yang, Vivek Kwatra, George Toderici

EarthVision: Large Scale Computer Vision for Remote Sensing Imagery
Invited Speaker: John Quinn

LatinX in Computer Vision Research
Organizer: Ruben Villegas

Fine-Grained Visual Categorization
Organizer: Kimberly Wilber

The Art of Robustness: Devil and Angel in Adversarial Machine Learning
Organizer: Florian Tramèr
Invited Speaker: Nicholas Carlini

AI for Content Creation
Organizers: Deqing Sun, Huiwen Chang, Lu Jiang
Invited Speaker: Chitwan Saharia

LOng-form VidEo Understanding
Invited Speaker: Cordelia Schmid

Visual Perception and Learning in an Open World
Invited Speaker: Rahul Sukthankar

Media Forensics
Organizer : Christoph Bregler
Technical Committee Members: Shruti Agarwal, Scott McCloskey, Peng Zhou

Vision Datasets Understanding
Organizer: José Lezama

Embedded Vision
Invited Speaker: Matthias Grundmann

Federated Learning for Computer Vision
Invited Speaker: Zheng Xu

Large Scale Holistic Video Understanding
Organizer: David Ross
Invited Speaker: Anurag Arnab

Learning With Limited Labelled Data for Image and Video Understanding
Invited Speaker: Hugo Larochelle

Bridging the Gap Between Computational Photography and Visual Recognition
Invited Speaker: Xiaohua Zhai

Explainable Artificial Intelligence for Computer Vision
Invited Speaker: Been Kim

Robustness in Sequential Data
Organizers: Sayna Ebrahimi, Kevin Murphy
Invited Speakers: Sayna Ebrahimi, Balaji Lakshminarayanan

Sketch-Oriented Deep Learning
Organizer: David Ha
Invited Speaker: Jonas Jongejan

Multimodal Learning and Applications
Invited Speaker: Cordelia Schmid

Computational Cameras and Displays
Organizer: Tali Dekel
Invited Speaker: Peyman Millanfar

Artificial Social Intelligence
Invited Speaker: Natasha Jaques

VizWiz Grand Challenge: Algorithms to Assist People Who Are Blind
Invited Speaker and Panelist: Andrew Howard

Image Matching: Local Features & Beyond
Organizer: Eduard Trulls

Multi-Agent Behavior: Representation, Modeling, Measurement, and Applications
Organizer: Ting Liu

Efficient Deep Learning for Computer Vision
Organizers: Pete Warden, Andrew Howard, Grace Chu, Jaeyoun Kim

Gaze Estimation and Prediction in the Wild
Organizer: Thabo Beeler

Tutorials

Denoising Diffusion-Based Generative Modeling: Foundations and Applications
Invited Speaker: Ruiqi Gao

Algorithmic Fairness: Why It’s Hard and Why It’s Interesting
Invited Speaker: Sanmi Koyejo

Beyond Convolutional Neural Networks
Invited Speakers: Neil Houlsby, Alexander Kolesnikov, Xiaohua Zhai

Joint Ego4D and Egocentric Perception, Interaction & Computing
Invited Speaker: Vittorio Ferrari

Deep AUC Maximization
Invited Speakers: Tianbao Yang

Vision-Based Robot Learning
Organizers: Michael S. Ryoo, Andy Zeng, Pete Florence

Graph Machine Learning for Visual Computing
Organizers: Federico Tombari
Invited Speakers: Federico Tombari, Fabian Manhardt



*Work done while at Google.  

Read More

Scanned Objects by Google Research: A Dataset of 3D-Scanned Common Household Items

Many recent advances in computer vision and robotics rely on deep learning, but training deep learning models requires a wide variety of data to generalize to new scenarios. Historically, deep learning for computer vision has relied on datasets with millions of items that were gathered by web scraping, examples of which include ImageNet, Open Images, YouTube-8M, and COCO. However, the process of creating these datasets can be labor-intensive, and can still exhibit labeling errors that can distort the perception of progress. Furthermore, this strategy does not readily generalize to arbitrary three-dimensional shapes or real-world robotic data.

Real-world robotic data collection is very useful, but difficult to scale and challenging to label.

Simulating robots and environments using tools such as Gazebo, MuJoCo, and Unity can mitigate many of the inherent limitations in these datasets. However, simulation is only an approximation of reality — handcrafted models built from polygons and primitives often correspond poorly to real objects. Even if a scene is built directly from a 3D scan of a real environment, the movable objects in that scan will act like fixed background scenery and will not respond the way real-world objects would. Due to these challenges, there are few large libraries with high-quality models of 3D objects that can be incorporated into physical and visual simulations to provide the variety needed for deep learning.

In “Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items”, presented at ICRA 2022, we describe our efforts to address this need by creating the Scanned Objects dataset, a curated collection of over 1000 3D-scanned common household items. The Scanned Objects dataset is usable in tools that read Simulation Description Format (SDF) models, including the Gazebo and PyBullet robotics simulators. Scanned Objects is hosted on Open Robotics, an open-source hosting environment for models compatible with the Gazebo simulator.

History
Robotics researchers within Google began scanning objects in 2011, creating high-fidelity 3D models of common household objects to help robots recognize and grasp things in their environments. However, it became apparent that 3D models have many uses beyond object recognition and robotic grasping, including scene construction for physical simulations and 3D object visualization for end-user applications. Therefore, this Scanned Objects project was expanded to bring 3D experiences to Google at scale, collecting a large number of 3D scans of household objects through a process that is more efficient and cost effective than traditional commercial-grade product photography.

Scanned Objects was an end-to-end effort, involving innovations at nearly every stage of the process, including curation of objects at scale for 3D scanning, the development of novel 3D scanning hardware, efficient 3D scanning software, fast 3D rendering software for quality assurance, and specialized frontends for web and mobile viewers. We also executed human-computer interaction studies to create effective experiences for interacting with 3D objects.

Objects that were acquired for scanning.

These object models proved useful in 3D visualizations for Everyday Robots, which used the models to bridge the sim-to-real gap for training, work later published as RetinaGAN and RL-CycleGAN. Building on these earlier 3D scanning efforts, in 2019 we began preparing an external version of the Scanned Objects dataset and transforming the previous set of 3D images into graspable 3D models.

Object Scanning
To create high-quality models, we built a scanning rig to capture images of an object from multiple directions under controlled and carefully calibrated conditions. The system consists of two machine vision cameras for shape detection, a DSLR camera for high-quality HDR color frame extraction, and a computer-controlled projector for pattern recognition. The scanning rig uses a structured light technique that infers a 3D shape from camera images with patterns of light that are projected onto an object.

The scanning rig used to capture 3D models.
A shoe being scanned (left). Images are captured from several directions with different patterns of light and color. A shadow passing over an object (right) illustrates how a 3D shape can be captured with an off-axis view of a shadow edge.

Simulation Model Conversion
The early internal scanned models used protocol buffer metadata, high-resolution visuals, and formats that were not suitable for simulation. For some objects, physical properties, such as mass, were captured by weighing the objects at scanning time, but surface properties, such as friction or deformation, were not represented.

So, following data collection, we built an automated pipeline to solve these issues and enable the use of scanned models in simulation systems. The automated pipeline filters out invalid or duplicate objects, automatically assigns object names using text descriptions of the objects, and eliminates object mesh scans that do not meet simulation requirements. Next, the pipeline estimates simulation properties (e.g., mass and moment of inertia) from shape and volume, constructs collision volumes, and downscales the model to a usable size. Finally, the pipeline converts each model to SDF format, creates thumbnail images, and packages the model for use in simulation systems.

The pipeline filters models that are not suitable for simulation, generates collision volumes, computes physical properties, downsamples meshes, generates thumbnails, and packages them all for use in simulation systems.
A collection of Scanned Object models rendered in Blender.

The output of this pipeline is a simulation model in an appropriate format with a name, mass, friction, inertia, and collision information, along with searchable metadata in a public interface compatible with our open-source hosting on Open Robotics’ Gazebo.

The output objects are represented as SDF models that refer to Wavefront OBJ meshes averaging 1.4 Mb per model. Textures for these models are in PNG format and average 11.2 Mb. Together, these provide high resolution shape and texture.

Impact
The Scanned Objects dataset contains 1030 scanned objects and their associated metadata, totaling 13 Gb, licensed under the CC-BY 4.0 License. Because these models are scanned rather than modeled by hand, they realistically reflect real object properties, not idealized recreations, reducing the difficulty of transferring learning from simulation to the real world.

Input views (left) and reconstructed shape and texture from two novel views on the right (figure from Differentiable Stereopsis).
Visualized action scoring predictions over three real-world 3D scans from the Replica dataset and Scanned Objects (figure from Where2Act).

The Scanned Objects dataset has already been used in over 25 papers across as many projects, spanning computer vision, computer graphics, robot manipulation, robot navigation, and 3D shape processing. Most projects used the dataset to provide synthetic training data for learning algorithms. For example, the Scanned Objects dataset was used in Kubric, an open-sourced generator of scalable datasets for use in over a dozen vision tasks, and in LAX-RAY, a system for searching shelves with lateral access X-rays to automate the mechanical search for occluded objects on shelves.

Unsupervised 3D keypoints on real-world data (figure from KeypointDeformer).

We hope that the Scanned Objects dataset will be used by more robotics and simulation researchers in the future, and that the example set by this dataset will inspire other owners of 3D model repositories to make them available for researchers everywhere. If you would like to try it yourself, head to Gazebo and start browsing!

Acknowledgments
The authors thank the Scanned Objects team, including Peter Anderson-Sprecher, J.J. Blumenkranz, James Bruce, Ken Conley, Katie Dektar, Charles DuHadway, Anthony Francis, Chaitanya Gharpure, Topraj Gurung, Kristy Headley, Ryan Hickman, John Isidoro, Sumit Jain, Brandon Kinman, Greg Kline, Mach Kobayashi, Nate Koenig, Kai Kohlhoff, James Kuffner, Thor Lewis, Mike Licitra, Lexi Martin, Julian (Mac) Mason, Rus Maxham, Pascal Muetschard, Kannan Pashupathy, Barbara Petit, Arshan Poursohi, Jared Russell, Matt Seegmiller, John Sheu, Joe Taylor, Josh Weaver, and Tommy McHugh.

Special thanks go to Krista Reyman for organizing this project, helping write the paper, and editing this blogpost, James Bruce for the scanning pipeline design and Pascal Muetschard for maintaining the database of object models.

Read More

LIMoE: Learning Multiple Modalities with One Sparse Mixture of Experts Model

Sparse models stand out among the most promising approaches for the future of deep learning. Instead of every part of a model processing every input (“dense” modeling), sparse models employing conditional computation learn to route individual inputs to different “experts” in a potentially huge network. This has many benefits. First, model size can increase while keeping computational cost constant — an effective and environmentally friendlier way to scale models, which is often key to high performance. Sparsity also naturally compartmentalizes neural networks. Dense models that learn many different tasks simultaneously (multitask) or sequentially (continual learning) often suffer negative interference, where too much task variety means it is better to just train one model per task, or catastrophic forgetting, where the model becomes worse at earlier tasks as new ones are added. Sparse models help avoid both these phenomena — by not applying the whole model to all inputs, “experts” in the model can specialize on different tasks or data types while still taking advantage of shared parts of the model.

Research on sparsity has long been pursued at Google Research. Pathways summarizes the research vision of building one single large model that diligently handles thousands of tasks and numerous data modalities. So far there has been considerable progress in sparse unimodal models for language (Switch, Task-MoE, GLaM) and computer vision (Vision MoE). Today, we take another important step towards the Pathways vision by studying large sparse models that simultaneously handle images and text with modality-agnostic routing. A relevant approach is multimodal contrastive learning, which requires a solid understanding of both images and text in order to align pictures with their correct text description. The strongest models that tackle this task to date rely on independent networks for each modality (a “two-tower” approach).

In “Multimodal Contrastive Learning with LIMoE: the Language Image Mixture of Experts”, we present the first large-scale multimodal architecture using a sparse mixture of experts. It simultaneously processes both images and text, but uses sparsely activated experts that naturally specialize. On zero-shot image classification, LIMoE outperforms both comparable dense multimodal models and two-tower approaches. The largest LIMoE achieves 84.1% zero-shot ImageNet accuracy, comparable to more expensive state-of-the-art models. Sparsity enables LIMoE to scale up gracefully and learn to handle very different inputs, addressing the tension between being a jack-of-all-trades generalist and a master-of-one specialist.

The LIMoE architecture contains many “experts” and routers decide which tokens (parts of an image or sentence) go to which experts. After being processed by expert layers (gray) and shared dense layers (brown), a final output layer computes a single vector representation for either an image or a text.

Sparse Mixture of Expert Models
Transformers represent data as a sequence of vectors (or tokens). Though originally developed for text, they can be applied to most things that are representable as a sequence of tokens, e.g., images, videos, and audio. Recent large-scale MoE models add expert layers to the Transformer architecture (e.g., gShard and ST-MoE in natural language processing, and Vision MoE for vision tasks).

A standard Transformer consists of many “blocks”, each containing various different layers. One of these layers is a feed-forward network (FFN). For LIMoE and the works cited above, this single FFN is replaced by an expert layer that contains many parallel FFNs, each of which is an expert. Given a sequence of tokens to process, a simple router learns to predict which experts should handle which tokens. Only a small number of experts are activated per token, meaning although the model capacity is significantly increased by virtue of having so many experts, the actual computational cost is controlled by using them sparsely. If only one expert is activated, the model’s cost is roughly equivalent to the standard Transformer model.

LIMoE does precisely that, activating one expert per example, thereby matching the computational cost of the dense baselines. What’s different is that the LIMoE router might see tokens of either image or text data.

A unique failure mode of MoE models occurs when they try to send all tokens to the same expert. Typically this is addressed with auxiliary losses, extra training objectives that encourage balanced expert usage. We found that dealing with multiple modalities interacted with sparsity to cause new failure modes that existing auxiliary losses could not address. To overcome this, we developed new auxiliary losses (more details in the paper) and used routing prioritization (BPR) during training, two innovations that resulted in stable and high performance multimodal models.

The new auxiliary losses (LIMoE aux) and routing prioritization (BPR) stabilized and improved overall performance (left) and increased the success rate of routing behavior (middle and right). A low success rate means the router does not use all the experts available and drops many tokens due to individual expert capacity being reached, which usually indicates the sparse model is not learning well. The combination introduced for LIMoE ensures high routing success rates for both images and text and consequently leads to significantly better performance.

Contrastive Learning with LIMoE
In multimodal contrastive learning, models are trained on paired image-text data (e.g., a photo and its caption). Typically, an image model extracts a representation of images, and a different text model extracts a representation of text. The contrastive learning objective encourages the image and text representations to be close for the same image-text pair and far away for content from different pairs. Such models with aligned representations can be adapted to new tasks without extra training data (“zero-shot”), e.g., an image will be classified as a dog if its representation is closer to the representation of the word “dog” than the word “cat”. This idea scales to thousands of classes and is referred to as zero-shot image classification.

CLIP and ALIGN (both two-tower models) scaled this process to achieve 76.2% and 76.4% zero-shot classification accuracy on the popular ImageNet dataset. We study one-tower models which compute both image and text representations. We find this reduces performance for dense models, likely due to negative interference or insufficient capacity. However, a compute-matched LIMoE not only improves over the one-tower dense model, but also outperforms two-tower dense models. We trained a series of models in a comparable training regimen to CLIP. Our dense L/16 model achieves 73.5% zero-shot accuracy, whereas LIMoE-L/16 gets to 78.6%, even outperforming CLIP’s more expensive, two-tower L/14 model (76.2%). As shown below, LIMoE’s use of sparsity provides a remarkable performance boost over dense models with equivalent cost.

For a given compute cost (x-axis), LIMoE models (circles, solid line) are significantly better than their dense baselines (triangles, dashed line). The architecture indicates the size of the underlying transformer, increasing from left (S/32) to right (L/16). Following standard convention, S (small), B (base), and L (large) refer to model scale. The number refers to the patch size, where smaller patches imply a larger architecture.

LiT and BASIC pushed zero-shot accuracy for dense two-tower models to 84.5% and 85.6% respectively. In addition to scaling, these approaches made use of specialized pre-training methods, repurposing image models that were already of exceptionally high quality. LIMoE-H/14 does not benefit from any pre-training or modality-specific components, but still achieved a comparable 84.1% zero-shot accuracy training from scratch. The scale of these models is also interesting to compare: LiT and BASIC are 2.1B and 3B parameter models. LIMoE-H/14 has 5.6B parameters in total, but via sparsity it only applies 675M parameters per token making it significantly more lightweight.

Data seen during training
Model   Pre-training     Image-text     Total      Parameters per token     ImageNet accuracy  
CLIP 12.8B 12.8B ~200M 76.2%
ALIGN 19.8B 19.8B ~410M 76.4%
LiT 25.8B 18.2B 44.0B 1.1B 84.5%
BASIC 19.7B 32.8B 52.5B 1.5B 85.6%
LIMoE H/14    23.3B 23.3B 675M 84.1%

Understanding LIMoE’s Behavior
LIMoE was motivated by the intuition that sparse conditional computation enables a generalist multimodal model to still develop the specialization needed to excel at understanding each modality. We analyzed LIMoE’s expert layers and uncovered a few interesting phenomena.

First, we see the emergence of modality-specialized experts. In our training setup there are many more image tokens than text tokens, so all experts tend to process at least some images, but some experts process either mostly images, mostly text, or both.

Distributions for an eight expert LIMoE; percentages indicate the amount of image tokens processed by the expert. There are one or two experts clearly specialized on text (shown by the mostly blue experts), usually two to four image specialists (mostly red), and the remainder are somewhere in the middle.

There are also some clear qualitative patterns among the image experts — e.g., in most LIMoE models, there is an expert that processes all image patches that contain text. In the example below, one expert processes fauna and greenery, and another processes human hands.

LIMoE chooses an expert for each token. Here we show which image tokens go to which experts on one of the layers of LIMoE-H/14. Despite not being trained to do so, we observe the emergence of semantic experts that specialize in specific topics such as plants or wheels.

Moving Forward
Multimodal models that handle many tasks are a promising route forward, and there are two key ingredients for success: scale, and the ability to avoid interference between distinct tasks and modalities while taking advantage of synergies. Sparse conditional computation is an excellent way of doing both. It enables performant and efficient generalist models that also have the capacity and flexibility for the specialization necessary to excel at individual tasks, as demonstrated by LIMoE’s solid performance with less compute.

Acknowledgements
We thank our co-authors on this work: Joan Puigcerver, Rodolphe Jenatton and Neil Houlsby. We also thank Andreas Steiner, Xiao Wang and Xiaohua Zhai, who led early explorations into dense single-tower models for contrastive multimodal learning, and also were instrumental in providing data access. We enjoyed useful discussions with André Susano Pinto, Maxim Neumann, Barret Zoph, Liam Fedus, Wei Han and Josip Djolonga. Finally, we would also like to thank and acknowledge Tom Small for the awesome animated figure used in this post.

Read More

Building a more helpful browser with machine learning

At Google we use technologies like machine learning (ML) to build more useful products — from filtering out email spam, to keeping maps up to date, to offering more relevant search results. Chrome is no exception: We use ML to make web images more accessible to people who are blind or have low vision, and we also generate real-time captions for online videos, in service of people in noisy environments, and those who are hard of hearing.

This work in Chrome continues, so we wanted to share some recent and future ML improvements that offer a safer, more accessible and more personalized browsing experience. Importantly: these updates are powered by on-device ML models, which means your data stays private, and never leaves your device.

More peace of mind, less annoying prompts

Safe Browsing in Chrome helps protect billions of devices every day, by showing warnings when people try to navigate to dangerous sites or download dangerous files (see the big red example below). Starting in March of this year, we rolled out a new ML model that identifies 2.5 times more potentially malicious sites and phishing attacks as the previous model – resulting in a safer and more secure web.

To further improve the browsing experience, we’re also evolving how people interact with web notifications. On the one hand, page notifications help deliver updates from sites you care about; on the other hand, notification permission prompts can become a nuisance. To help people browse the web with minimal interruption, Chrome predicts when permission prompts are unlikely to be granted based on how the user previously interacted with similar permission prompts, and silences these undesired prompts. In the next release of Chrome, we’re launching an ML model that makes these predictions entirely on-device.

Two separate images side by side. The first on the left is a smartphone showing a red screen and a warning message about phishing. The image on the right shows a Chrome browser window showing a pop-up message saying “Notifications blocked”.

With the next release of Chrome, this is what you will see if a phishing attempt is detected (Left) and Chrome will show permission requests quietly when the user is unlikely to grant them (Right).

Finding what’s important, always in your language

Earlier this year we launched Journeys to help people retrace their steps online. For example: You might spend weeks planning a national park visit – researching attractions, comparing flights and shopping for gear. With ML and Journeys, Chrome brings together the pages you’ve visited about a given topic, and makes it easy to pick up where you left off (vs. scr o o o l l ling through your browser history).

When you return to those hiking boots and camping guides, we’re also using ML to make those websites available in your preferred language. In particular, we’ve launched an updated language identification model to figure out the language of the page, and whether it needs to be translated to match your preferences. As a result, we’re seeing tens of millions more successful translations every day.

A Chrome browser showing Journeys related to travel. The user can see a cluster of recent searches they did related to a trip to Yosemite.

The Journeys feature of Chrome groups together your search history based on topic or intent.

A browser built just for you

Maybe you like to read news articles in the morning – phone in one hand, cereal spoon in the other – so you share lots of links from Chrome. Or maybe voice search is more your thing, as you sneak in a few questions during your transit ride to work. Either way, we want to make sure Chrome is meeting you where you’re at, so in the near future, we’ll be using ML to adjust the toolbar in real-time – highlighting the action that’s most useful in that moment (e.g., share link, voice search, etc.). Of course, you’ll be able to customize it manually as well.

A Chrome browser with a highlighted square around an icon to the right of the address bar. At the top is a share icon, and at the bottom is a microphone icon.

The toolbar in Chrome on Android will adapt based on your needs.

Our goal is to build a browser that’s genuinely and continuously helpful, and we’re excited about the possibilities that ML provides. At the end of the day, though, your experience is what really matters, so please tweet @googlechrome to send us your feedback.

Read More

End-to-end Generative Pre-training for Multimodal Video Captioning

Multimodal video captioning systems utilize both the video frames and speech to generate natural language descriptions (captions) of videos. Such systems are stepping stones towards the longstanding goal of building multimodal conversational systems that effortlessly communicate with users while perceiving environments through multimodal input streams.

Unlike video understanding tasks (e.g., video classification and retrieval) where the key challenge lies in processing and understanding multimodal input videos, the task of multimodal video captioning includes the additional challenge of generating grounded captions. The most widely adopted approach for this task is to train an encoder-decoder network jointly using manually annotated data. However, due to a lack of large-scale, manually annotated data, the task of annotating grounded captions for videos is labor intensive and, in many cases, impractical. Previous research such as VideoBERT and CoMVT pre-train their models on unlabelled videos by leveraging automatic speech recognition (ASR). However, such models often cannot generate natural language sentences because they lack a decoder, and thus only the video encoder is transferred to the downstream tasks.

In “End-to-End Generative Pre-training for Multimodal Video Captioning”, published at CVPR 2022, we introduce a novel pre-training framework for multimodal video captioning. This framework, which we call multimodal video generative pre-training or MV-GPT, jointly trains a multimodal video encoder and a sentence decoder from unlabelled videos by leveraging a future utterance as the target text and formulating a novel bi-directional generation task. We demonstrate that MV-GPT effectively transfers to multimodal video captioning, achieving state-of-the-art results on various benchmarks. Additionally, the multimodal video encoder is competitive for multiple video understanding tasks, such as VideoQA, text-video retrieval, and action recognition.

Future Utterance as an Additional Text Signal
Typically, each training video clip for multimodal video captioning is associated with two different texts: (1) a speech transcript that is aligned with the clip as a part of the multimodal input stream, and (2) a target caption, which is often manually annotated. The encoder learns to fuse information from the transcript with visual contents, and the target caption is used to train the decoder for generation. However, in the case of unlabelled videos, each video clip comes only with a transcript from ASR, without a manually annotated target caption. Moreover, we cannot use the same text (the ASR transcript) for the encoder input and decoder target, since the generation of the target would then be trivial.

MV-GPT circumvents this challenge by leveraging a future utterance as an additional text signal and enabling joint pre-training of the encoder and decoder. However, training a model to generate future utterances that are often not grounded in the input content is not ideal. So we apply a novel bi-directional generation loss to reinforce the connection to the input.

Bi-directional Generation Loss
The issue of non-grounded text generation is mitigated by formulating a bi-directional generation loss that includes forward and backward generation. Forward generation produces future utterances given visual frames and their corresponding transcripts and allows the model to learn to fuse the visual content with its corresponding transcript. Backward generation takes the visual frames and future utterances to train the model to generate a transcript that contains more grounded text of the video clip. Bi-directional generation loss in MV-GPT allows the encoder and the decoder to be trained to handle visually grounded texts.

Bi-directional generation in MV-GPT. A model is trained with two generation losses. In forward generation, the model generates a future utterance (blue boxes) given the frames and the present utterance (red boxes), whereas the present is generated from the future utterance in backward generation. Two special beginning-of-sentence tokens ([BOS-F] and [BOS-B]) initiate forward and backward generation for the decoder.

Results on Multimodal Video Captioning
We compare MV-GPT to existing pre-training losses using the same model architecture, on YouCook2 with standard evaluation metrics (Bleu-4, Cider, Meteor and Rouge-L). While all pre-training techniques improve captioning performances, it is critical to pre-train the decoder jointly to improve model performance. We demonstrate that MV-GPT outperforms the previous state-of-the-art joint pre-training method by over 3.5% with relative gains across all four metrics.

Pre-training Loss Pre-trained Parts Bleu-4 Cider Meteor Rouge-L
No Pre-training N/A 13.25 1.03 17.56 35.48
CoMVT Encoder 14.46 1.24 18.46 37.17
UniVL Encoder + Decoder 19.95 1.98 25.27 46.81
MV-GPT (ours) Encoder + Decoder 21.26 2.14 26.36 48.58
MV-GPT performance across four metrics (Bleu-4, Cider, Meteor and Rouge-L) of different pre-training losses on YouCook2. “Pre-trained parts” indicates which parts of the model are pre-trained — only the encoder or both the encoder and decoder. We reimplement the loss functions of existing methods but use our model and training strategies for a fair comparison.

We transfer a model pre-trained by MV-GPT to four different captioning benchmarks: YouCook2, MSR-VTT, ViTT and ActivityNet-Captions. Our model achieves state-of-the-art performance on all four benchmarks by significant margins. For instance on the Meteor metric, MV-GPT shows over 12% relative improvements in all four benchmarks.

YouCook2 MSR-VTT ViTT ActivityNet-Captions
Best Baseline 22.35 29.90 11.00 10.90
MV-GPT (ours) 27.09 38.66 26.75 12.31
Meteor metric scores of the best baseline methods and MV-GPT on four benchmarks.

Results on Non-generative Video Understanding Tasks
Although MV-GPT is designed to train a generative model for multimodal video captioning, we also find that our pre-training technique learns a powerful multimodal video encoder that can be applied to multiple video understanding tasks, including VideoQA, text-video retrieval and action classification. When compared to the best comparable baseline models, the model transferred from MV-GPT shows superior performance in five video understanding benchmarks on their primary metrics — i.e., top-1 accuracy for VideoQA and action classification benchmarks, and recall at 1 for the retrieval benchmark.

Task Benchmark Best Comparable Baseline MV-GPT
VideoQA MSRVTT-QA 41.5 41.7
ActivityNet-QA 38.9 39.1
Text-Video Retrieval MSR-VTT 33.7 37.3
Action Recognition Kinetics-400 78.9 80.4
Kinetics-600 80.6 82.4
Comparisons of MV-GPT to best comparable baseline models on five video understanding benchmarks. For each dataset we report the widely used primary metric, i.e., MSRVTT-QA and ActivityNet-QA: Top-1 answer accuracy; MSR-VTT: Recall at 1; and Kinetics: Top-1 classification accuracy.

Summary
We introduce MV-GPT, a new generative pre-training framework for multimodal video captioning. Our bi-directional generative objective jointly pre-trains a multimodal encoder and a caption decoder by using utterances sampled at different times in unlabelled videos. Our pre-trained model achieves state-of-the-art results on multiple video captioning benchmarks and other video understanding tasks, namely VideoQA, video retrieval and action classification.

Acknowledgements
This research was conducted by Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab and Cordelia Schmid.

Read More

Deep Learning with Label Differential Privacy

Over the last several years, there has been an increased focus on developing differential privacy (DP) machine learning (ML) algorithms. DP has been the basis of several practical deployments in industry — and has even been employed by the U.S. Census — because it enables the understanding of system and algorithm privacy guarantees. The underlying assumption of DP is that changing a single user’s contribution to an algorithm should not significantly change its output distribution.

In the standard supervised learning setting, a model is trained to make a prediction of the label for each input given a training set of example pairs {[input1,label1], …, [inputn, labeln]}. In the case of deep learning, previous work introduced a DP training framework, DP-SGD, that was integrated into TensorFlow and PyTorch. DP-SGD protects the privacy of each example pair [input, label] by adding noise to the stochastic gradient descent (SGD) training algorithm. Yet despite extensive efforts, in most cases, the accuracy of models trained with DP-SGD remains significantly lower than that of non-private models.

DP algorithms include a privacy budget, ε, which quantifies the worst-case privacy loss for each user. Specifically, ε reflects how much the probability of any particular output of a DP algorithm can change if one replaces any example of the training set with an arbitrarily different one. So, a smaller ε corresponds to better privacy, as the algorithm is more indifferent to changes of a single example. However, since smaller ε tends to hurt model utility more, it is not uncommon to consider ε up to 8 in deep learning applications. Notably, for the widely used multiclass image classification dataset, CIFAR-10, the highest reported accuracy (without pre-training) for DP models with ε = 3 is 69.3%, a result that relies on handcrafted visual features. In contrast, non-private scenarios (ε = ∞) with learned features have shown to achieve >95% accuracy while using modern neural network architectures. This performance gap remains a roadblock for many real-world applications to adopt DP. Moreover, despite recent advances, DP-SGD often comes with increased computation and memory overhead due to slower convergence and the need to compute the norm of the per-example gradient.

In “Deep Learning with Label Differential Privacy”, presented at NeurIPS 2021, we consider a more relaxed, but important, special case called label differential privacy (LabelDP), where we assume the inputs (input1, …, inputn) are public, and only the privacy of the training labels (label1, …, labeln) needs to be protected. With this relaxed guarantee, we can design novel algorithms that utilize a prior understanding of the labels to improve the model utility. We demonstrate that LabelDP achieves 20% higher accuracy than DP-SGD on the CIFAR-10 dataset. Our results across multiple tasks confirm that LabelDP could significantly narrow the performance gap between private models and their non-private counterparts, mitigating the challenges in real world applications. We also present a multi-stage algorithm for training deep neural networks with LabelDP. Finally, we are excited to release the code for this multi-stage training algorithm.

LabelDP
The notion of LabelDP has been studied in the Probably Approximately Correct (PAC) learning setting, and captures several practical scenarios. Examples include: (i) computational advertising, where impressions are known to the advertiser and thus considered non-sensitive, but conversions reveal user interest and are thus private; (ii) recommendation systems, where the choices are known to a streaming service provider, but the user ratings are considered sensitive; and (iii) user surveys and analytics, where demographic information (e.g., age, gender) is non-sensitive, but income is sensitive.

We make several key observations in this scenario. (i) When only the labels need to be protected, much simpler algorithms can be applied for data preprocessing to achieve LabelDP without any modifications to the existing deep learning training pipeline. For example, the classic Randomized Response (RR) algorithm, designed to eliminate evasive answer biases in survey aggregation, achieves LabelDP by simply flipping the label to a random one with a probability that depends on ε. (ii) Conditioned on the (public) input, we can compute a prior probability distribution, which provides a prior belief of the likelihood of the class labels for the given input. With a novel variant of RR, RR-with-prior, we can incorporate prior information to reduce the label noise while maintaining the same privacy guarantee as classical RR.

The figure below illustrates how RR-with-prior works. Assume a model is built to classify an input image into 10 categories. Consider a training example with the label “airplane”. To guarantee LabelDP, classical RR returns a random label sampled according to a given distribution (see the top-right panel of the figure below). The smaller the targeted privacy budget ε is, the larger the probability of sampling an incorrect label has to be. Now assume we have a prior probability showing that the given input is “likely an object that flies” (lower left panel). With the prior, RR-with-prior will discard all labels with small prior and only sample from the remaining labels. By dropping these unlikely labels, the probability of returning the correct label is significantly increased, while maintaining the same privacy budget ε (lower right panel).

Randomized response: If no prior information is given (top-left), all classes are sampled with equal probability. The probability of sampling the true class (P[airplane] ≈ 0.5) is higher if the privacy budget is higher (top-right). RR-with-prior: Assuming a prior distribution (bottom-left), unlikely classes are “suppressed” from the sampling distribution (bottom-right). So the probability of sampling the true class (P[airplane] ≈ 0.9) is increased under the same privacy budget.

A Multi-stage Training Algorithm
Based on the RR-with-prior observations, we present a multi-stage algorithm for training deep neural networks with LabelDP. First, the training set is randomly partitioned into multiple subsets. An initial model is then trained on the first subset using classical RR. Finally, the algorithm divides the data into multiple parts, and at each stage, a single part is used to train the model. The labels are produced using RR-with-prior, and the priors are based on the prediction of the model trained so far.

An illustration of the multi-stage training algorithm. The training set is partitioned into t disjoint subsets. An initial model is trained on the first subset using classical RR. Then the trained model is used to provide prior predictions in the RR-with-prior step and in the training of the later stages.

Results
We benchmark the multi-stage training algorithm’s empirical performance on multiple datasets, domains, and architectures. On the CIFAR-10 multi-class classification task for the same privacy budget ε, the multi-stage training algorithm (blue in the figure below) guaranteeing LabelDP achieves 20% higher accuracy than DP-SGD. We emphasize that LabelDP protects only the labels while DP-SGD protects both the inputs and labels, so this is not a strictly fair comparison. Nonetheless, this result demonstrates that for specific application scenarios where only the labels need to be protected, LabelDP could lead to significant improvements in the model utility while narrowing the performance gap between private models and public baselines.

Comparison of the model utility (test accuracy) of different algorithms under different privacy budgets.

In some domains, prior knowledge is naturally available or can be built using publicly available data only. For example, many machine learning systems have historical models which could be evaluated on new data to provide label priors. In domains where unsupervised or self-supervised learning algorithms work well, priors could also be built from models pre-trained on unlabeled (therefore public with respect to LabelDP) data. Specifically, we demonstrate two self-supervised learning algorithms in our CIFAR-10 evaluation (orange and green traces in the figure above). We use self-supervised learning models to compute representations for the training examples and run k-means clustering on the representations. Then, we spend a small amount of privacy budget (ε ≤ 0.05) to query a histogram of the label distribution of each cluster and use that as the label prior for the points in each cluster. This prior significantly boosts the model utility in the low privacy budget regime (ε < 1).

Similar observations hold across multiple datasets such as MNIST, Fashion-MNIST and non-vision domains, such as the MovieLens-1M movie rating task. Please see our paper for the full report on the empirical results.

The empirical results suggest that protecting the privacy of the labels can be significantly easier than protecting the privacy of both the inputs and labels. This can also be mathematically proven under specific settings. In particular, we can show that for convex stochastic optimization, the sample complexity of algorithms privatizing the labels is much smaller than that of algorithms privatizing both labels and inputs. In other words, to achieve the same level of model utility under the same privacy budget, LabelDP requires fewer training examples.

Conclusion
We demonstrated that both empirical and theoretical results suggest that LabelDP is a promising relaxation of the full DP guarantee. In applications where the privacy of the inputs does not need to be protected, LabelDP could reduce the performance gap between a private model and the non-private baseline. For future work, we plan to design better LabelDP algorithms for other tasks beyond multi-class classification. We hope that the release of the multi-stage training algorithm code provides researchers with a useful resource for DP research.

Acknowledgements
This work was carried out in collaboration with Badih Ghazi, Noah Golowich, and Ravi Kumar. We also thank Sami Torbey for valuable feedback on our work.

Read More

Image-Text Pre-training with Contrastive Captioners

Oftentimes, machine learning (ML) model developers begin their design using a generic backbone model that is trained at scale and with capabilities transferable to a wide range of downstream tasks. In natural language processing, a number of popular backbone models, including BERT, T5, GPT-3 (sometimes also referred to as “foundation models”), are pre-trained on web-scale data and have demonstrated generic multi-tasking capabilities through zero-shot, few-shot or transfer learning. Compared with training over-specialized individual models, pre-training backbone models for a large number of downstream tasks can amortize the training costs, allowing one to overcome resource limitations when building large scale models.

In computer vision, pioneering work has shown the effectiveness of single-encoder models pre-trained for image classification to capture generic visual representations that are effective for other downstream tasks. More recently, contrastive dual-encoder (CLIP, ALIGN, Florence) and generative encoder-decoder (SimVLM) approaches trained using web-scale noisy image-text pairs have been explored. Dual-encoder models exhibit remarkable zero-shot image classification capabilities but are less effective for joint vision-language understanding. On the other hand, encoder-decoder methods are good at image captioning and visual question answering but cannot perform retrieval-style tasks.

In “CoCa: Contrastive Captioners are Image-Text Foundation Models”, we present a unified vision backbone model called Contrastive Captioner (CoCa). Our model is a novel encoder-decoder approach that simultaneously produces aligned unimodal image and text embeddings and joint multimodal representations, making it flexible enough to be directly applicable for all types of downstream tasks. Specifically, CoCa achieves state-of-the-art results on a series of vision and vision-language tasks spanning vision recognition, cross-modal alignment, and multimodal understanding. Furthermore, it learns highly generic representations so that it can perform as well or better than fully fine-tuned models with zero-shot learning or frozen encoders.

Overview of Contrastive Captioners (CoCa) compared to single-encoder, dual-encoder and encoder-decoder models.

Method
We propose CoCa, a unified training framework that combines contrastive loss and captioning loss on a single training data stream consisting of image annotations and noisy image-text pairs, effectively merging single-encoder, dual-encoder and encoder-decoder paradigms.

To this end, we present a novel encoder-decoder architecture where the encoder is a vision transformer (ViT), and the text decoder transformer is decoupled into two parts, a unimodal text decoder and a multimodal text decoder. We skip cross-attention in unimodal decoder layers to encode text-only representations for contrastive loss, and cascade multimodal decoder layers with cross-attention to image encoder outputs to learn multimodal image-text representations for captioning loss. This design maximizes the model’s flexibility and universality in accommodating a wide spectrum of tasks, and at the same time, it can be efficiently trained with a single forward and backward propagation for both training objectives, resulting in minimal computational overhead. Thus, the model can be trained end-to-end from scratch with training costs comparable to a naïve encoder-decoder model.

Illustration of forward propagation used by CoCa for both contrastive and captioning losses.

Benchmark Results
The CoCa model can be directly fine-tuned on many tasks with minimal adaptation. By doing so, our model achieves a series of state-of-the-art results on popular vision and multimodal benchmarks, including (1) visual recognition: ImageNet, Kinetics-400/600/700, and MiT; (2) cross-modal alignment: MS-COCO, Flickr30K, and MSR-VTT; and (3) multimodal understanding: VQA, SNLI-VE, NLVR2, and NoCaps.

Comparison of CoCa with other image-text backbone models (without task-specific customization) and multiple state-of-the-art task-specialized models.

It is noteworthy that CoCa attains these results as a single model adapted for all tasks while often lighter than prior top-performing specialized models. For example, CoCa obtains 91.0% ImageNet top-1 accuracy while using less than half the parameters of prior state-of-the-art models. In addition, CoCa also obtains strong generative capability of high-quality image captions.

Image classification scaling performance comparing fine-tuned ImageNet top-1 accuracy versus model size.
Text captions generated by CoCa with NoCaps images as input.

Zero-Shot Performance
Besides achieving excellent performance with fine-tuning, CoCa also outperforms previous state-of-the-art models on zero-shot learning tasks, including image classification,and cross-modal retrieval. CoCa obtains 86.3% zero-shot accuracy on ImageNet while also robustly outperforming prior models on challenging variant benchmarks, such as ImageNet-A, ImageNet-R, ImageNet-V2, and ImageNet-Sketch. As shown in the figure below, CoCa obtains better zero-shot accuracy with smaller model sizes compared to prior methods.

Image classification scaling performance comparing zero-shot ImageNet top-1 accuracy versus model size.

Frozen Encoder Representation
One particularly exciting observation is that CoCa achieves results comparable to the best fine-tuned models using only a frozen visual encoder, in which features extracted after model training are used to train a classifier, rather than the more computationally intensive effort of fine-tuning a model. On ImageNet, a frozen CoCa encoder with a learned classification head obtains 90.6% top-1 accuracy, which is better than the fully fine-tuned performance of existing backbone models (90.1%). We also find this setup to work extremely well for video recognition. We feed sampled video frames into the CoCa frozen image encoder individually, and fuse output features by attentional pooling before applying a learned classifier. This simple approach using a CoCa frozen image encoder achieves video action recognition top-1 accuracy of 88.0% on Kinetics-400 dataset and demonstrates that CoCa learns a highly generic visual representation with the combined training objectives.

Comparison of Frozen CoCa visual encoder with (multiple) best-performing fine-tuned models.

Conclusion
We present Contrastive Captioner (CoCa), a novel pre-training paradigm for image-text backbone models. This simple method is widely applicable to many types of vision and vision-language downstream tasks, and obtains state-of-the-art performance with minimal or even no task-specific adaptations.

Acknowledgements
We would like to thank our co-authors Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu who have been involved in all aspects of the project. We also would like to thank Yi-Ting Chen, Kaifeng Chen, Ye Xia, Zhen Li, Chao Jia, Yinfei Yang, Zhengdong Zhang, Wei Han, Yuan Cao, Tao Zhu, Futang Peng, Soham Ghosh, Zihang Dai, Xin Li, Anelia Angelova, Jason Baldridge, Izhak Shafran, Shengyang Dai, Abhijit Ogale, Zhifeng Chen, Claire Cui, Paul Natsev, Tom Duerig for helpful discussions, Andrew Dai for help with contrastive models, Christopher Fifty and Bowen Zhang for help with video models, Yuanzhong Xu for help with model scaling, Lucas Beyer for help with data preparation, Andy Zeng for help with MSR-VTT evaluation, Hieu Pham and Simon Kornblith for help with zero-shot evaluations, Erica Moreira and Victor Gomes for help with resource coordination, Liangliang Cao for proofreading, Tom Small for creating the animations used in this blogpost, and others in the Google Brain team for support throughout this project.

Read More

How we build with and for people with disabilities

Editor’s note: Today is Global Accessibility Awareness Day. We’re also sharing how we’re making education more accessibleand launching a newAndroid accessibility feature.

Over the past nine years, my job has focused on building accessible products and supporting Googlers with disabilities. Along the way, I’ve been constantly reminded of how vast and diverse the disability community is, and how important it is to continue working alongside this community to build technology and solutions that are truly helpful.

Before delving into some of the accessibility features our teams have been building, I want to share how we’re working to be more inclusive of people with disabilities to create more accessible tools overall.

Nothing about us, without us

In the disability community, people often say “nothing about us without us.” It’s a sentiment that I find sums up what disability inclusion means. The types of barriers that people with disabilities face in society vary depending on who they are, where they live and what resources they have access to. No one’s experience is universal. That’s why it’s essential to include a wide array of people with disabilities at every stage of the development process for any of our accessibility products, initiatives or programs.

We need to work to make sure our teams at Google are reflective of the people we’re building for. To do so, last year we launched our hiring site geared toward people with disabilities — including our Autism Career Program to further grow and strengthen our autistic community. Most recently, we helped launch the Neurodiversity Career Connector along with other companies to create a job portal that connects neurodiverse candidates to companies that are committed to hiring more inclusively.

Beyond our internal communities, we also must partner with communities outside of Google so we can learn what is truly useful to different groups and parlay that understanding into the improvement of current products or the creation of new ones. Those partnerships have resulted in the creation of Project Relate, a communication tool for people with speech impairments, the development of a completely new TalkBack, Android’s built-in screen reader, and the improvement of Select-to-Speak, a Chromebook tool that lets you hear selected text on your screen spoken out loud.

Equitable experiences for everyone

Engaging and listening to these communities — inside and outside of Google — make it possible to create tools and features like the ones we’re sharing today.

The ability to add alt-text, which is a short description of an image that is read aloud by screen readers, directly to images sent through Gmail starts rolling out today. With this update, people who use screen readers will know what’s being sent to them, whether it’s a GIF celebrating the end of the week or a screenshot of an important graph.

Communication tools that are inclusive of everyone are especially important as teams have shifted to fully virtual or hybrid meetings. Again, everyone experiences these changes differently. We’ve heard from some people who are deaf or hard of hearing, that this shift has made it easier to identify who is speaking — something that is often more difficult in person. But, in the case of people who use ASL, we’ve heard that it can be difficult to be in a virtual meeting and simultaneously see their interpreter and the person speaking to them.

Multi-pin, a new feature in Google Meet, helps solve this. Now you can pin multiple video tiles at once, for example, the presenter’s screen and the interpreter’s screen. And like many accessibility features, the usefulness extends beyond people with disabilities. The next time someone is watching a panel and wants to pin multiple people to the screen, this feature makes that possible.

We’ve also been working to make video content more accessible to those who are blind or low-vision through audio descriptions that describe verbally what is on the screen visually. All of our English language YouTube Originals content from the past year — and moving forward — will now have English audio descriptions available globally. To turn on the audio description track, at the bottom right of the video player, click on “Settings”, select “Audio track”, and choose “English descriptive”.

For many people with speech impairments, being understood by the technology that powers tools like voice typing or virtual assistants can be difficult. In 2019, we started work to change that through Project Euphonia, a research initiative that works with community organizations and people with speech impairments to create more inclusive speech recognition models. Today, we’re expanding Project Euphonia’s research to include four more languages: French, Hindi, Japanese and Spanish. With this expansion, we can create even more helpful technology for more people — no matter where they are or what language they speak.

I’ve learned so much in my time working in this space and among the things I’ve learned is the absolute importance of building right alongside the very people who will most use these tools in the end. We’ll continue to do that as we work to create a more inclusive and accessible world, both physically and digitally.

Read More

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 8×8 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 256×256. 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
Rate
FID IS
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 256×256. 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.

Conclusion
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.

Acknowledgements
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.

Read More

Contextual Rephrasing in Google Assistant

When people converse with one another, context and references play a critical role in driving their conversation more efficiently. For instance, if one asks the question “Who wrote Romeo and Juliet?” and, after receiving an answer, asks “Where was he born?”, it is clear that ‘he’ is referring to William Shakespeare without the need to explicitly mention him. Or if someone mentions “python” in a sentence, one can use the context from the conversation to determine whether they are referring to a type of snake or a computer language. If a virtual assistant cannot robustly handle context and references, users would be required to adapt to the limitation of the technology by repeating previously shared contextual information in their follow-up queries to ensure that the assistant understands their requests and can provide relevant answers.

In this post, we present a technology currently deployed on Google Assistant that allows users to speak in a natural manner when referencing context that was defined in previous queries and answers. The technology, based on the latest machine learning (ML) advances, rephrases a user’s follow-up query to explicitly mention the missing contextual information, thus enabling it to be answered as a stand-alone query. While Assistant considers many types of context for interpreting the user input, in this post we are focusing on short-term conversation history.

Context Handling by Rephrasing
One of the approaches taken by Assistant to understand contextual queries is to detect if an input utterance is referring to previous context and then rephrase it internally to explicitly include the missing information. Following on from the previous example in which the user asked who wrote Romeo and Juliet, one may ask follow-up questions like “When?”. Assistant recognizes that this question is referring to both the subject (Romeo and Juliet) and answer from the previous query (William Shakespeare) and can rephrase “When?” to “When did William Shakespeare write Romeo and Juliet?”

While there are other ways to handle context, for instance, by applying rules directly to symbolic representations of the meaning of queries, like intents and arguments, the advantage of the rephrasing approach is that it operates horizontally at the string level across any query answering, parsing, or action fulfillment module.

Conversation on a smart display device, where Assistant understands multiple contextual follow-up queries, allowing the user to have a more natural conversation. The phrases appearing at the bottom of the display are suggestions for follow-up questions that the user can select. However, the user can still ask different questions.

A Wide Variety of Contextual Queries
The natural language processing field, traditionally, has not put much emphasis on a general approach to context, focusing on the understanding of stand-alone queries that are fully specified. Accurately incorporating context is a challenging problem, especially when considering the large variety of contextual query types. The table below contains example conversations that illustrate query variability and some of the many contextual challenges that Assistant’s rephrasing method can resolve (e.g., differentiating between referential and non-referential cases or identifying what context a query is referencing). We demonstrate how Assistant is now able to rephrase follow-up queries, adding contextual information before providing an answer.

System Architecture
At a high level, the rephrasing system generates rephrasing candidates by using different types of candidate generators. Each rephrasing candidate is then scored based on a number of signals, and the one with the highest score is selected.

High level architecture of Google Assistant contextual rephraser.

Candidate Generation
To generate rephrasing candidates we use a hybrid approach that applies different techniques, which we classify into three categories:

  1. Generators based on the analysis of the linguistic structure of the queries use grammatical and morphological rules to perform specific operations — for instance, the replacement of pronouns or other types of referential phrases with antecedents from the context.
  2. Generators based on query statistics combine key terms from the current query and its context to create candidates that match popular queries from historical data or common query patterns.
  3. Generators based on Transformer technologies, such as MUM, learn to generate sequences of words according to a number of training samples. LaserTagger and FELIX are technologies suitable for tasks with high overlap between the input and output texts, are very fast at inference time, and are not vulnerable to hallucination (i.e., generating text that is not related to the input texts). Once presented with a query and its context, they can generate a sequence of text edits to transform the input queries into a rephrasing candidate by indicating which portions of the context should be preserved and which words should be modified.

Candidate Scoring
We extract a number of signals for each rephrasing candidate and use an ML model to select the most promising candidate. Some of the signals depend only on the current query and its context. For example, is the topic of the current query similar to the topic of the previous query? Or, is the current query a good stand-alone query or does it look incomplete? Other signals depend on the candidate itself: How much of the information of the context does the candidate preserve? Is the candidate well-formed from a linguistic point of view? Etc.

Recently, new signals generated by BERT and MUM models have significantly improved the performance of the ranker, fixing about one-third of the recall headroom while minimizing false positives on query sequences that are not contextual (and therefore do not require a rephrasing).

Example conversation on a phone where Assistant understands a sequence of contextual queries.

Conclusion
The solution described here attempts to resolve contextual queries by rephrasing them in order to make them fully answerable in a stand-alone manner, i.e., without having to relate to other information during the fulfillment phase. The benefit of this approach is that it is agnostic to the mechanisms that would fulfill the query, thus making it usable as a horizontal layer to be deployed before any further processing.

Given the variety of contexts naturally used in human languages, we adopted a hybrid approach that combines linguistic rules, large amounts of historic data through logs, and ML models based on state-of-the-art Transformer approaches. By generating a number of rephrasing candidates for each query and its context, and then scoring and ranking them using a variety of signals, Assistant can rephrase and thus correctly interpret most contextual queries. As Assistant can handle most types of linguistic references, we are empowering users to have more natural conversations. To make such multi-turn conversations even less cumbersome, Assistant users can turn on Continued Conversation mode to enable asking follow-up queries without the need to repeat “Hey Google” between each query. We are also using this technology in other virtual assistant settings, for instance, interpreting context from something shown on a screen or playing on a speaker.

Acknowledgements
This post reflects the combined work of Aliaksei Severyn, André Farias, Cheng-Chun Lee, Florian Thöle, Gabriel Carvajal, Gyorgy Gyepesi, Julien Cretin, Liana Marinescu, Martin Bölle, Patrick Siegler, Sebastian Krause, Victor Ähdel, Victoria Fossum, Vincent Zhao. We also thank Amar Subramanya, Dave Orr, Yury Pinsky for helpful discussions and support.

Read More