Timezone: »
Previous work has proposed many new loss functions and regularizers that improve test accuracy on image classification tasks. However, it is not clear whether these loss functions learn better representations for downstream tasks. This paper studies how the choice of training objective affects the transferability of the hidden representations of convolutional neural networks trained on ImageNet. We show that many objectives lead to statistically significant improvements in ImageNet accuracy over vanilla softmax cross-entropy, but the resulting fixed feature extractors transfer substantially worse to downstream tasks, and the choice of loss has little effect when networks are fully fine-tuned on the new tasks. Using centered kernel alignment to measure similarity between hidden representations of networks, we find that differences among loss functions are apparent only in the last few layers of the network. We delve deeper into representations of the penultimate layer, finding that different objectives and hyperparameter combinations lead to dramatically different levels of class separation. Representations with higher class separation obtain higher accuracy on the original task, but their features are less useful for downstream tasks. Our results suggest there exists a trade-off between learning invariant features for the original task and features relevant for transfer tasks.
Author Information
Simon Kornblith (Google Brain)
Ting Chen (Google Brain)
Honglak Lee (U. Michigan)
Mohammad Norouzi (Google Brain)
More from the Same Authors
-
2021 : Palette: Image-to-Image Diffusion Models »
Chitwan Saharia · William Chan · Huiwen Chang · Chris Lee · Jonathan Ho · Tim Salimans · David Fleet · Mohammad Norouzi -
2021 : Understanding and Improving Robustness of VisionTransformers through patch-based NegativeAugmentation »
Yao Qin · Chiyuan Zhang · Ting Chen · Balaji Lakshminarayanan · Alex Beutel · Xuezhi Wang -
2021 : Learning Action Translator for Meta Reinforcement Learning on Sparse-Reward Tasks »
Yijie Guo · Qiucheng Wu · Honglak Lee -
2021 : Fast Inference and Transfer of Compositional Task for Few-shot Task Generalization »
Sungryull Sohn · Hyunjae Woo · Jongwook Choi · Izzeddin Gur · Aleksandra Faust · Honglak Lee -
2021 : Learning Parameterized Task Structure for Generalization to Unseen Entities »
Anthony Liu · Sungryull Sohn · Honglak Lee -
2021 : SURF: Semi-supervised Reward Learning with Data Augmentation for Feedback-efficient Preference-based Reinforcement Learning »
Jongjin Park · Younggyo Seo · Jinwoo Shin · Honglak Lee · Pieter Abbeel · Kimin Lee -
2021 : Learning compositional tasks from language instructions »
Lajanugen Logeswaran · Wilka Carvalho · Honglak Lee -
2021 : Palette: Image-to-Image Diffusion Models »
Chitwan Saharia · William Chan · Huiwen Chang · Chris Lee · Jonathan Ho · Tim Salimans · David Fleet · Mohammad Norouzi -
2022 : Allele-conditional attention mechanism for HLA-peptide complex binding affinity prediction »
Rodrigo Hormazabal · Doyeong Hwang · Kiyoung Kim · Sehui Han · Kyunghoon Bae · Honglak Lee -
2022 : Dynamics-Augmented Decision Transformer for Offline Dynamics Generalization »
Changyeon Kim · Junsu Kim · Younggyo Seo · Kimin Lee · Honglak Lee · Jinwoo Shin -
2022 : Human alignment of neural network representations »
Lukas Muttenthaler · Lorenz Linhardt · Jonas Dippel · Robert Vandermeulen · Simon Kornblith -
2022 : Learning Exploration Policies with View-based Intrinsic Rewards »
Yijie Guo · Yao Fu · Run Peng · Honglak Lee -
2022 : Invited Talk: Mohammad Norouzi »
Mohammad Norouzi -
2022 : Interactive Industrial Panel »
Jiahao Sun · Ahmed Ibrahim · Marjan Ghazvininejad · Yu Cheng · Boxing Chen · Mohammad Norouzi · Rahul Gupta -
2022 : ReSPack: A Large-Scale Rectilinear Steiner Tree Packing Data Generator and Benchmark »
Kanghoon Lee · Youngjoon Park · Han-Seul Jeong · Deunsol Yoon · Sunghoon Hong · Sungryull Sohn · Minu Kim · Hanbum Ko · Moontae Lee · Honglak Lee · Kyunghoon Kim · Euihyuk Kim · Seonggeon Cho · Jaesang Min · Woohyung Lim -
2022 Poster: Patching open-vocabulary models by interpolating weights »
Gabriel Ilharco · Mitchell Wortsman · Samir Yitzhak Gadre · Shuran Song · Hannaneh Hajishirzi · Simon Kornblith · Ali Farhadi · Ludwig Schmidt -
2022 Poster: Transferring Pre-trained Multimodal Representations with Cross-modal Similarity Matching »
Byoungjip Kim · Sungik Choi · Dasol Hwang · Moontae Lee · Honglak Lee -
2022 Poster: Pure Transformers are Powerful Graph Learners »
Jinwoo Kim · Dat Nguyen · Seonwoo Min · Sungjun Cho · Moontae Lee · Honglak Lee · Seunghoon Hong -
2022 Poster: OpenSRH: optimizing brain tumor surgery using intraoperative stimulated Raman histology »
Cheng Jiang · Asadur Chowdury · Xinhai Hou · Akhil Kondepudi · Christian Freudiger · Kyle Conway · Sandra Camelo-Piragua · Daniel Orringer · Honglak Lee · Todd Hollon -
2022 Poster: Understanding and Improving Robustness of Vision Transformers through Patch-based Negative Augmentation »
Yao Qin · Chiyuan Zhang · Ting Chen · Balaji Lakshminarayanan · Alex Beutel · Xuezhi Wang -
2022 Poster: Transformers meet Stochastic Block Models: Attention with Data-Adaptive Sparsity and Cost »
Sungjun Cho · Seonwoo Min · Jinwoo Kim · Moontae Lee · Honglak Lee · Seunghoon Hong -
2022 Poster: A Unified Sequence Interface for Vision Tasks »
Ting Chen · Saurabh Saxena · Lala Li · Tsung-Yi Lin · David Fleet · Geoffrey Hinton -
2022 Poster: UniCLIP: Unified Framework for Contrastive Language-Image Pre-training »
Janghyeon Lee · Jongsuk Kim · Hyounguk Shon · Bumsoo Kim · Seung Hwan Kim · Honglak Lee · Junmo Kim -
2022 Poster: CEDe: A collection of expert-curated datasets with atom-level entity annotations for Optical Chemical Structure Recognition »
Rodrigo Hormazabal · Changyoung Park · Soonyoung Lee · Sehui Han · Yeonsik Jo · Jaewan Lee · Ahra Jo · Seung Hwan Kim · Jaegul Choo · Moontae Lee · Honglak Lee -
2022 Expo Talk Panel: Towards learning agents for solving complex real-world tasks »
Honglak Lee -
2021 Poster: Improving Contrastive Learning on Imbalanced Data via Open-World Sampling »
Ziyu Jiang · Tianlong Chen · Ting Chen · Zhangyang Wang -
2021 Poster: Improving Transferability of Representations via Augmentation-Aware Self-Supervision »
Hankook Lee · Kibok Lee · Kimin Lee · Honglak Lee · Jinwoo Shin -
2021 Poster: Intriguing Properties of Contrastive Losses »
Ting Chen · Calvin Luo · Lala Li -
2021 Poster: Successor Feature Landmarks for Long-Horizon Goal-Conditioned Reinforcement Learning »
Christopher Hoang · Sungryull Sohn · Jongwook Choi · Wilka Carvalho · Honglak Lee -
2021 Poster: Generalized Shape Metrics on Neural Representations »
Alex H Williams · Erin Kunz · Simon Kornblith · Scott Linderman -
2021 Poster: Meta-learning to Improve Pre-training »
Aniruddh Raghu · Jonathan Lorraine · Simon Kornblith · Matthew McDermott · David Duvenaud -
2021 Poster: Improved Transformer for High-Resolution GANs »
Long Zhao · Zizhao Zhang · Ting Chen · Dimitris Metaxas · Han Zhang -
2021 Poster: Environment Generation for Zero-Shot Compositional Reinforcement Learning »
Izzeddin Gur · Natasha Jaques · Yingjie Miao · Jongwook Choi · Manoj Tiwari · Honglak Lee · Aleksandra Faust -
2021 Poster: Do Vision Transformers See Like Convolutional Neural Networks? »
Maithra Raghu · Thomas Unterthiner · Simon Kornblith · Chiyuan Zhang · Alexey Dosovitskiy -
2020 Poster: Memory Based Trajectory-conditioned Policies for Learning from Sparse Rewards »
Yijie Guo · Jongwook Choi · Marcin Moczulski · Shengyu Feng · Samy Bengio · Mohammad Norouzi · Honglak Lee -
2020 Poster: The Origins and Prevalence of Texture Bias in Convolutional Neural Networks »
Katherine L. Hermann · Ting Chen · Simon Kornblith -
2020 Oral: The Origins and Prevalence of Texture Bias in Convolutional Neural Networks »
Katherine L. Hermann · Ting Chen · Simon Kornblith -
2020 Poster: Big Self-Supervised Models are Strong Semi-Supervised Learners »
Ting Chen · Simon Kornblith · Kevin Swersky · Mohammad Norouzi · Geoffrey E Hinton -
2020 : Policy Panel »
Roya Pakzad · Dia Kayyali · Marzyeh Ghassemi · Shakir Mohamed · Mohammad Norouzi · Ted Pedersen · Anver Emon · Abubakar Abid · Darren Byler · Samhaa R. El-Beltagy · Nayel Shafei · Mona Diab -
2020 Affinity Workshop: Muslims in ML »
Marzyeh Ghassemi · Mohammad Norouzi · Shakir Mohamed · Aya Salama · Tasmie Sarker -
2020 Poster: Bridging Imagination and Reality for Model-Based Deep Reinforcement Learning »
Guangxiang Zhu · Minghao Zhang · Honglak Lee · Chongjie Zhang -
2019 : Poster Session »
Ethan Harris · Tom White · Oh Hyeon Choung · Takashi Shinozaki · Dipan Pal · Katherine L. Hermann · Judy Borowski · Camilo Fosco · Chaz Firestone · Vijay Veerabadran · Benjamin Lahner · Chaitanya Ryali · Fenil Doshi · Pulkit Singh · Sharon Zhou · Michel Besserve · Michael Chang · Anelise Newman · Mahesan Niranjan · Jonathon Hare · Daniela Mihai · Marios Savvides · Simon Kornblith · Christina M Funke · Aude Oliva · Virginia de Sa · Dmitry Krotov · Colin Conwell · George Alvarez · Alex Kolchinski · Shengjia Zhao · Mitchell Gordon · Michael Bernstein · Stefano Ermon · Arash Mehrjou · Bernhard Schölkopf · John Co-Reyes · Michael Janner · Jiajun Wu · Josh Tenenbaum · Sergey Levine · Yalda Mohsenzadeh · Zhenglong Zhou -
2019 Poster: When does label smoothing help? »
Rafael Müller · Simon Kornblith · Geoffrey E Hinton -
2019 Spotlight: When does label smoothing help? »
Rafael Müller · Simon Kornblith · Geoffrey E Hinton -
2019 Poster: Saccader: Improving Accuracy of Hard Attention Models for Vision »
Gamaleldin Elsayed · Simon Kornblith · Quoc V Le -
2019 Poster: High Fidelity Video Prediction with Large Stochastic Recurrent Neural Networks »
Ruben Villegas · Arkanath Pathak · Harini Kannan · Dumitru Erhan · Quoc V Le · Honglak Lee -
2018 Poster: A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks »
Kimin Lee · Kibok Lee · Honglak Lee · Jinwoo Shin -
2018 Spotlight: A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks »
Kimin Lee · Kibok Lee · Honglak Lee · Jinwoo Shin -
2018 Poster: Hierarchical Reinforcement Learning for Zero-shot Generalization with Subtask Dependencies »
Sungryull Sohn · Junhyuk Oh · Honglak Lee -
2018 Poster: Learning Hierarchical Semantic Image Manipulation through Structured Representations »
Seunghoon Hong · Xinchen Yan · Thomas Huang · Honglak Lee -
2017 : Invited Talk 5 »
Honglak Lee -
2017 Workshop: Learning Disentangled Features: from Perception to Control »
Emily Denton · Siddharth Narayanaswamy · Tejas Kulkarni · Honglak Lee · Diane Bouchacourt · Josh Tenenbaum · David Pfau -
2017 Poster: Bridging the Gap Between Value and Policy Based Reinforcement Learning »
Ofir Nachum · Mohammad Norouzi · Kelvin Xu · Dale Schuurmans -
2017 Poster: Filtering Variational Objectives »
Chris Maddison · John Lawson · George Tucker · Nicolas Heess · Mohammad Norouzi · Andriy Mnih · Arnaud Doucet · Yee Teh -
2017 Poster: Value Prediction Network »
Junhyuk Oh · Satinder Singh · Honglak Lee -
2016 Poster: Perspective Transformer Nets: Learning Single-View 3D Object Reconstruction without 3D Supervision »
Xinchen Yan · Jimei Yang · Ersin Yumer · Yijie Guo · Honglak Lee -
2016 Poster: Learning What and Where to Draw »
Scott E Reed · Zeynep Akata · Santosh Mohan · Samuel Tenka · Bernt Schiele · Honglak Lee -
2016 Oral: Learning What and Where to Draw »
Scott E Reed · Zeynep Akata · Santosh Mohan · Samuel Tenka · Bernt Schiele · Honglak Lee -
2015 : Deep Learning for Real-Time Atari Game Play Using Offline Monte-Carlo Tree Search Planning »
Honglak Lee -
2015 Symposium: Deep Learning Symposium »
Yoshua Bengio · Marc'Aurelio Ranzato · Honglak Lee · Max Welling · Andrew Y Ng -
2015 Poster: Deep Visual Analogy-Making »
Scott E Reed · Yi Zhang · Yuting Zhang · Honglak Lee -
2015 Poster: Action-Conditional Video Prediction using Deep Networks in Atari Games »
Junhyuk Oh · Xiaoxiao Guo · Honglak Lee · Richard L Lewis · Satinder Singh -
2015 Spotlight: Action-Conditional Video Prediction using Deep Networks in Atari Games »
Junhyuk Oh · Xiaoxiao Guo · Honglak Lee · Richard L Lewis · Satinder Singh -
2015 Oral: Deep Visual Analogy-Making »
Scott E Reed · Yi Zhang · Yuting Zhang · Honglak Lee -
2015 Poster: Learning Structured Output Representation using Deep Conditional Generative Models »
Kihyuk Sohn · Honglak Lee · Xinchen Yan -
2015 Poster: Weakly-supervised Disentangling with Recurrent Transformations for 3D View Synthesis »
Jimei Yang · Scott E Reed · Ming-Hsuan Yang · Honglak Lee -
2014 Workshop: Representation and Learning Methods for Complex Outputs »
Richard Zemel · Dale Schuurmans · Kilian Q Weinberger · Yuhong Guo · Jia Deng · Francesco Dinuzzo · Hal Daumé III · Honglak Lee · Noah A Smith · Richard Sutton · Jiaqian YU · Vitaly Kuznetsov · Luke Vilnis · Hanchen Xiong · Calvin Murdock · Thomas Unterthiner · Jean-Francis Roy · Martin Renqiang Min · Hichem SAHBI · Fabio Massimo Zanzotto -
2014 Poster: Deep Learning for Real-Time Atari Game Play Using Offline Monte-Carlo Tree Search Planning »
Xiaoxiao Guo · Satinder Singh · Honglak Lee · Richard L Lewis · Xiaoshi Wang -
2014 Poster: Improved Multimodal Deep Learning with Variation of Information »
Kihyuk Sohn · Wenling Shang · Honglak Lee -
2013 Poster: Robust Image Denoising with Multi-Column Deep Neural Networks »
Forest Agostinelli · Michael R Anderson · Honglak Lee -
2012 Poster: Learning to Align from Scratch »
Gary B Huang · Marwan A Mattar · Honglak Lee · Erik Learned-Miller -
2010 Workshop: Deep Learning and Unsupervised Feature Learning »
Honglak Lee · Marc'Aurelio Ranzato · Yoshua Bengio · Geoffrey E Hinton · Yann LeCun · Andrew Y Ng -
2009 Poster: Unsupervised feature learning for audio classification using convolutional deep belief networks »
Honglak Lee · Peter Pham · Yan Largman · Andrew Y Ng -
2007 Poster: Sparse deep belief net model for visual area V2 »
Honglak Lee · Ekanadham Chaitanya · Andrew Y Ng -
2006 Poster: Efficient sparse coding algorithms, end-stopping and nCRF surround suppression »
Honglak Lee · Alexis Battle · Raina Rajat · Andrew Y Ng