Attention-based Iterative Decomposition for Tensor Product Representation

TL;DR

This paper introduces an Attention-based Iterative Decomposition (AID) module that improves the extraction of symbolic structure from data in Tensor Product Representation models, enhancing their systematic generalization capabilities.

Contribution

The paper proposes a novel AID module that enhances decomposition in TPR models, leading to better systematic generalization and more accurate structural representations.

Findings

AID significantly improves TPR-based model performance on generalization tasks.

AID produces more compositional and well-bound structural representations.

Experimental results demonstrate the effectiveness of AID in various evaluations.

Abstract

In recent research, Tensor Product Representation (TPR) is applied for the systematic generalization task of deep neural networks by learning the compositional structure of data. However, such prior works show limited performance in discovering and representing the symbolic structure from unseen test data because their decomposition to the structural representations was incomplete. In this work, we propose an Attention-based Iterative Decomposition (AID) module designed to enhance the decomposition operations for the structured representations encoded from the sequential input data with TPR. Our AID can be easily adapted to any TPR-based model and provides enhanced systematic decomposition through a competitive attention mechanism between input features and structured representations. In our experiments, AID shows effectiveness by significantly improving the performance of TPR-based prior works on the series of systematic generalization tasks. Moreover, in the quantitative and qualitative evaluations, AID produces more compositional and well-bound structural representations than other works.