Poor Extrapolation Performance

Poor extrapolation performance, the inability of machine learning models to generalize predictions beyond the range of their training data, is a significant challenge across various fields, including reinforcement learning, natural language processing, and scientific machine learning. Current research focuses on mitigating this issue through techniques like data synthesis, regularization methods targeting Q-functions or behavior policies, and the development of more robust model architectures such as physics-informed neural networks. Addressing this limitation is crucial for deploying machine learning models reliably in real-world applications where encountering out-of-distribution data is inevitable, and for enhancing the trustworthiness and generalizability of scientific discoveries driven by data-driven models.

Papers

July 17, 2024

Sparsity-based Safety Conservatism for Constrained Offline Reinforcement Learning
Minjae Cho, Chuangchuang Sun
Reinforcement Learning Extrapolation Framework Bi Level Optimization Poor Extrapolation Performance

October 20, 2023

Let's Synthesize Step by Step: Iterative Dataset Synthesis with Large Language Models by Extrapolating Errors from Small Models
Ruida Wang, Wangchunshu Zhou, Mrinmaya Sachan
Synthetic Dataset Data Synthesis Small Model Dataset Synthesis Poor Extrapolation Performance Validation Dataset

July 25, 2023

Offline Reinforcement Learning with On-Policy Q-Function Regularization
Laixi Shi, Robert Dadashi, Yuejie Chi, Pablo Samuel Castro, Matthieu Geist
Offline Reinforcement Learning Estimation Task Q Function Policy Regularization Poor Extrapolation Performance

June 15, 2023

Understanding and Mitigating Extrapolation Failures in Physics-Informed Neural Networks
Lukas Fesser, Luca D'Amico-Wong, Richard Qiu
Physic Informed Neural Network Human Understanding High Dimensional Partial Differential Equation PINN Method Extrapolation Performance Poor Extrapolation Performance

April 27, 2023

Learning to Extrapolate: A Transductive Approach
Aviv Netanyahu, Abhishek Gupta, Max Simchowitz, Kaiqing Zhang, Pulkit Agrawal
LeArning Abstract Reparameterization Method Transductive Approach Combinatorial Generalization Poor Extrapolation Performance

February 7, 2023

IB-UQ: Information bottleneck based uncertainty quantification for neural function regression and neural operator learning
Ling Guo, Hao Wu, Wenwen Zhou, Yan Wang, Tao Zhou
Uncertainty Quantification Neural Operator Information Bottleneck Deep Regression Confidence Map Uncertainty Quantification Method Poor Extrapolation Performance

December 16, 2022

Interpretable models for extrapolation in scientific machine learning
Eric S. Muckley, James E. Saal, Bryce Meredig, Christopher S. Roper, John H. Martin
Interpretable Model Model Performance Black Box Model Extrapolation Framework Scientific Machine Learning Trend Extrapolation Poor Extrapolation Performance

September 13, 2022

Revisiting Neural Scaling Laws in Language and Vision
Ibrahim Alabdulmohsin, Behnam Neyshabur, Xiaohua Zhai
Deep Learning Human Language Vision Paper Scaling Law Neural Scaling Law Poor Extrapolation Performance BIG Bench