/phyre-qa

Primary LanguageJupyter Notebook

Question Answering Network for Physical Reasoning

Introduction

This is the repository to evaluate the question-answering module for physical reasoning in PHYRE benchmark.
Our goal is to let agent learn the generalized physical representation through reasoning module. (Deep Learning, Korea University)

Method

Learned embedding from question-answer module is applied into the main network through FiLM-based attention.

Results

Models Cross Within
DQN 36.8 77.6
ATQAN 39.0 80.2

Our model was able to surpass the baseline through AUCESS score, in PHYRE-B task.

References

PHYRE