On a Built-in Conflict between Deep Learning and Systematic Generalization

TL;DR

This paper investigates a fundamental conflict in deep learning where internal function sharing hampers systematic generalization, revealing that common models tend to reuse decision boundaries, which limits their ability to generalize systematically.

Contribution

The study identifies internal function sharing as a key factor weakening systematic generalization across various deep learning architectures.

Findings

Function sharing leads to fewer decision boundaries for new outputs.

Standard models like CNNs, Transformers, and LSTMs exhibit this conflict.

Insights may guide future research to improve systematic generalization.

Abstract

In this paper, we hypothesize that internal function sharing is one of the reasons to weaken o.o.d. or systematic generalization in deep learning for classification tasks. Under equivalent prediction, a model partitions an input space into multiple parts separated by boundaries. The function sharing prefers to reuse boundaries, leading to fewer parts for new outputs, which conflicts with systematic generalization. We show such phenomena in standard deep learning models, such as fully connected, convolutional, residual networks, LSTMs, and (Vision) Transformers. We hope this study provides novel insights into systematic generalization and forms a basis for new research directions.