Timezone: »
Poster
Bidirectional Learning for Offline Infinite-width Model-based Optimization
Can Chen · Yingxueff Zhang · Jie Fu · Xue (Steve) Liu · Mark Coates
In offline model-based optimization, we strive to maximize a black-box objective function by only leveraging a static dataset of designs and their scores. This problem setting arises in numerous fields including the design of materials, robots, DNAs, proteins, etc. Recent approaches train a deep neural network (DNN) model on the static dataset to act as a proxy function, and then perform gradient ascent on the existing designs to obtain potentially high-scoring designs. This methodology frequently suffers from the out-of-distribution problem where the proxy function often returns adversarial designs. To mitigate this problem, we propose $\textit{\textbf{B}i\textbf{D}irectional learning for offline \textbf{I}nfinite-width model-based optimization}~(\textbf{BDI})$. BDI consists of two mappings: the forward mapping leverages the static dataset to predict the scores of the high-scoring designs, and the backward mapping leverages the high-scoring designs to predict the scores of the static dataset. The backward mapping, neglected in previous work, can distill more information of the static dataset into the high-scoring designs, which effectively mitigates the out-of-distribution problem. Yet, for a finite-width DNN model, the loss function of the backward mapping is intractable and only has an approximate form, which leads to a significant deterioration of the design quality. We thus adopt an infinite-width DNN model and propose to employ the corresponding neural tangent kernel to yield a closed-form loss for more accurate design updates. Experiments on various tasks verify the effectiveness of BDI. The code is available [here](https://github.com/GGchen1997/BDI).
Author Information
Can Chen (McGill University & MILA)
Yingxueff Zhang (Huawei Technology Canada)
Jie Fu (University of Montreal)
Xue (Steve) Liu (McGill University)
Mark Coates (McGill University)
More from the Same Authors
-
2021 : Calculus of Consent via MARL: Legitimating the Collaborative Governance Supplying Public Goods »
Yang Hu · Zhui Zhu · Sirui Song · Xue (Steve) Liu · Yang Yu -
2022 Poster: Versatile Multi-stage Graph Neural Network for Circuit Representation »
shuwen yang · Zhihao Yang · Dong Li · Yingxueff Zhang · Zhanguang Zhang · Guojie Song · Jianye Hao -
2022 : Structural Causal Model for Molecular Dynamics Simulation »
Qi Liu · Yuanqi Du · Fan Feng · Qiwei Ye · Jie Fu -
2022 : Graph neural networks for Ramsey graphs »
Amur Ghose · Amit Levi · Yingxueff Zhang -
2022 : Precise Augmentation and Counting of Helicobacter Pylori in Histology Image »
· Yixin Chen · Zhifeng Shuai · Fang Peng · Yanbo Lv · Luoning Zheng · Xue (Steve) Liu · Antoni Chan · Tei-Wei Kuo · Chun Jason XUE -
2022 Spotlight: Lightning Talks 3A-2 »
shuwen yang · Xu Zhang · Delvin Ce Zhang · Lan-Zhe Guo · Renzhe Xu · Zhuoer Xu · Yao-Xiang Ding · Weihan Li · Xingxuan Zhang · Xi-Zhu Wu · Zhenyuan Yuan · Hady Lauw · Yu Qi · Yi-Ge Zhang · Zhihao Yang · Guanghui Zhu · Dong Li · Changhua Meng · Kun Zhou · Gang Pan · Zhi-Fan Wu · Bo Li · Minghui Zhu · Zhi-Hua Zhou · Yafeng Zhang · Yingxueff Zhang · shiwen cui · Jie-Jing Shao · Zhanguang Zhang · Zhenzhe Ying · Xiaolong Chen · Yu-Feng Li · Guojie Song · Peng Cui · Weiqiang Wang · Ming GU · Jianye Hao · Yihua Huang -
2022 Spotlight: Versatile Multi-stage Graph Neural Network for Circuit Representation »
shuwen yang · Zhihao Yang · Dong Li · Yingxueff Zhang · Zhanguang Zhang · Guojie Song · Jianye Hao -
2021 Poster: Generalized DataWeighting via Class-Level Gradient Manipulation »
Can Chen · Shuhao Zheng · Xi Chen · Erqun Dong · Xue (Steve) Liu · Hao Liu · Dejing Dou