Explainable Neural Computation via Stack Neural Module Networks

TL;DR

This paper introduces a neural modular approach for compositional reasoning in complex tasks like question answering, which is interpretable and does not require supervised reasoning traces during training.

Contribution

The proposed model automatically induces sub-task decomposition and shares modules across tasks, enhancing interpretability without strong supervision.

Findings

Model is more interpretable to humans than existing methods.

Users can better understand the reasoning process and predict success or failure.

Experiments demonstrate improved interpretability over state-of-the-art models.

Abstract

In complex inferential tasks like question answering, machine learning models must confront two challenges: the need to implement a compositional reasoning process, and, in many applications, the need for this reasoning process to be interpretable to assist users in both development and prediction. Existing models designed to produce interpretable traces of their decision-making process typically require these traces to be supervised at training time. In this paper, we present a novel neural modular approach that performs compositional reasoning by automatically inducing a desired sub-task decomposition without relying on strong supervision. Our model allows linking different reasoning tasks though shared modules that handle common routines across tasks. Experiments show that the model is more interpretable to human evaluators compared to other state-of-the-art models: users can better understand the model's underlying reasoning procedure and predict when it will succeed or fail based on observing its intermediate outputs.