Attribution Preservation in Network Compression for Reliable Network Interpretation

TL;DR

This paper identifies a conflict between network compression and input attribution in neural networks, proposing a framework that preserves attribution quality during compression to ensure reliable interpretation in safety-critical applications.

Contribution

It introduces a novel attribution preservation framework using Weighted Collapsed Attribution Matching to maintain attribution fidelity during network compression.

Findings

The proposed method effectively preserves attribution maps across various compression techniques.

Preserving attributions improves the reliability of neural network interpretations post-compression.

The framework demonstrates both quantitative and qualitative improvements in attribution consistency.

Abstract

Neural networks embedded in safety-sensitive applications such as self-driving cars and wearable health monitors rely on two important techniques: input attribution for hindsight analysis and network compression to reduce its size for edge-computing. In this paper, we show that these seemingly unrelated techniques conflict with each other as network compression deforms the produced attributions, which could lead to dire consequences for mission-critical applications. This phenomenon arises due to the fact that conventional network compression methods only preserve the predictions of the network while ignoring the quality of the attributions. To combat the attribution inconsistency problem, we present a framework that can preserve the attributions while compressing a network. By employing the Weighted Collapsed Attribution Matching regularizer, we match the attribution maps of the network being compressed to its pre-compression former self. We demonstrate the effectiveness of our algorithm both quantitatively and qualitatively on diverse compression methods.