Timezone: »
Across applications spanning supervised classification and sequential control, deep learning has been reported to find shortcut'' solutions that fail catastrophically under minor changes in the data distribution. In this paper, we show empirically that DNNs can be coaxed to avoid poor shortcuts by providing an additional
priming'' feature computed from key input features, usually a coarse output estimate. Priming relies on approximate domain knowledge of these task-relevant key input features, which is often easy to obtain in practical settings. For example, one might prioritize recent frames over past frames in a video input for visual imitation learning, or salient foreground over background pixels for image classification. On NICO image classification, MuJoCo continuous control, and CARLA autonomous driving, our priming strategy works significantly better than several popular state-of-the-art approaches for feature selection and data augmentation. We connect these empirical findings to recent theoretical results on DNN optimization, and argue theoretically that priming distracts the optimizer away from poor shortcuts by creating better, simpler shortcuts.
Author Information
Chuan Wen (Tsinghua University)
Jianing Qian (University of Pennsylvania)
Jierui Lin (UT Austin)
Jiaye Teng (Tsinghua University)
Dinesh Jayaraman (University of Pennsylvania)
Yang Gao (Tsinghua University)
Related Events (a corresponding poster, oral, or spotlight)
-
2022 Spotlight: Fighting Fire with Fire: Avoiding DNN Shortcuts through Priming »
Thu. Jul 21st 02:35 -- 02:40 PM Room Ballroom 1 & 2
More from the Same Authors
-
2022 : Pre-Trained Image Encoder for Generalizable Visual Reinforcement Learning »
Zhecheng Yuan · Zhecheng Yuan · Zhengrong Xue · Zhengrong Xue · Bo Yuan · Bo Yuan · Xueqian Wang · Xueqian Wang · Yi Wu · Yi Wu · Yang Gao · Yang Gao · Huazhe Xu · Huazhe Xu -
2023 Poster: Finding Generalization Measures by Contrasting Signal and Noise »
Jiaye Teng · Bohang Zhang · Ruichen Li · Haowei He · Yequan Wang · Yan Tian · Yang Yuan -
2023 Poster: For Pre-Trained Vision Models in Motor Control, Not All Policy Learning Methods are Created Equal »
Yingdong Hu · Renhao Wang · Li Li · Yang Gao -
2023 Poster: Policy Contrastive Imitation Learning »
Jialei Huang · Zhao-Heng Yin · Yingdong Hu · Yang Gao -
2023 Poster: On Uni-Modal Feature Learning in Supervised Multi-Modal Learning »
Chenzhuang Du · Jiaye Teng · Tingle Li · Yichen Liu · Tianyuan Yuan · Yue Wang · Yang Yuan · Hang Zhao -
2023 Poster: LIV: Language-Image Representations and Rewards for Robotic Control »
Yecheng Jason Ma · Vikash Kumar · Amy Zhang · Osbert Bastani · Dinesh Jayaraman -
2022 Poster: Versatile Offline Imitation from Observations and Examples via Regularized State-Occupancy Matching »
Yecheng Jason Ma · Andrew Shen · Dinesh Jayaraman · Osbert Bastani -
2022 Spotlight: Versatile Offline Imitation from Observations and Examples via Regularized State-Occupancy Matching »
Yecheng Jason Ma · Andrew Shen · Dinesh Jayaraman · Osbert Bastani -
2021 Poster: T-SCI: A Two-Stage Conformal Inference Algorithm with Guaranteed Coverage for Cox-MLP »
Jiaye Teng · Zeren Tan · Yang Yuan -
2021 Spotlight: T-SCI: A Two-Stage Conformal Inference Algorithm with Guaranteed Coverage for Cox-MLP »
Jiaye Teng · Zeren Tan · Yang Yuan -
2021 Poster: Keyframe-Focused Visual Imitation Learning »
Chuan Wen · Jierui Lin · Jianing Qian · Yang Gao · Dinesh Jayaraman -
2021 Spotlight: Keyframe-Focused Visual Imitation Learning »
Chuan Wen · Jierui Lin · Jianing Qian · Yang Gao · Dinesh Jayaraman -
2020 Poster: Cautious Adaptation For Reinforcement Learning in Safety-Critical Settings »
Jesse Zhang · Brian Cheung · Chelsea Finn · Sergey Levine · Dinesh Jayaraman