Role Taxonomy of Units in Deep Neural Networks

TL;DR

This paper proposes a role taxonomy for units in deep neural networks using a retrieval-of-function test, linking unit roles to the network's generalization ability and providing insights into DNN mechanisms.

Contribution

It introduces a novel role taxonomy for DNN units and a retrieval-of-function test to categorize units based on their functional preferences, connecting unit roles to generalization.

Findings

Ratios of unit categories correlate with DNN generalization ability

The taxonomy reveals different functional roles of units in well-generalized DNNs

Provides signs of DNNs with strong generalization based on unit role distribution

Abstract

Identifying the role of network units in deep neural networks (DNNs) is critical in many aspects including giving understandings on the mechanisms of DNNs and building basic connections between deep learning and neuroscience. However, there remains unclear on which roles the units in DNNs with different generalization ability could present. To this end, we give role taxonomy of units in DNNs via introducing the retrieval-of-function test, where units are categorized into four types in terms of their functional preference on separately the training set and testing set. We show that ratios of the four categories are highly associated with the generalization ability of DNNs from two distinct perspectives, based on which we give signs of DNNs with well generalization.