Task and Explanation Network

TL;DR

The paper argues that explainability should be a core requirement for AI systems and introduces TENet, a two-head network that jointly performs a task and generates explanations. TENet uses a shared backbone to produce both a multi-label class output and explanatory words, optimized with a combined loss $L_{tot} = L_{task} + L_{expl}$, and makes predictions via top-$TOP_C$ classes and top-$TOP_W$ words. Preliminary experiments across multiple backbones on COCO show that RegNet extunderscore Y extunderscore 400MF achieves higher joint task and explanation accuracy than ResNet50, demonstrating the feasibility of integrated task-and-explanation models. This work highlights a path toward more interpretable AI systems by coupling decision outputs with human-readable explanations in a single architecture.

Abstract

Explainability in deep networks has gained increased importance in recent years. We argue herein that an AI must be tasked not just with a task but also with an explanation of why said task was accomplished as such. We present a basic framework -- Task and Explanation Network (TENet) -- which fully integrates task completion and its explanation. We believe that the field of AI as a whole should insist -- quite emphatically -- on explainability.

Figures (1)

• Figure 1: Four input images from the validation set, with the resultant TENet outputs.