Transparency by Design: Closing the Gap Between Performance and Interpretability in Visual Reasoning

Abstract

Visual question answering requires high-order reasoning about an image, which is a fundamental capability needed by machine systems to follow complex directives. Recently, modular networks have been shown to be an effective framework for performing visual reasoning tasks. While modular networks were initially designed with a degree of model transparency, their performance on complex visual reasoning benchmarks was lacking. Current stateof-the-art approaches do not provide an effective mechanism for understanding the reasoning process. In this paper, we close the performance gap between interpretable models and state-of-the-art visual reasoning methods. We propose a set of visual-reasoning primitives which, when composed, manifest as a model capable of performing complex reasoning tasks in an explicitly-interpretable manner. The fidelity and interpretability of the primitives' outputs enable an unparalleled ability to diagnose the strengths and weaknesses of the resulting model. Critically, we show that these primitives are highly performant, achieving state-of-the-art accuracy of 99.1% on the CLEVR dataset. We also show that our model is able to effectively learn generalized representations when provided a small amount of data containing novel object attributes. Using the CoGenT generalization task, we show more than a 20 percentage point improvement over the current state of the art.

CLEVR-Ref+: Diagnosing Visual Reasoning With Referring Expressions

“…Very interestingly, we found that the module trained on translating the last module output to segmentation mask is general, and can produce excellent humaninterpretable segmentation masks when attached to intermediate module outputs, revealing the entire reasoning process. We believe ours is the first to show clean visualization of the visual reasoning process carried out by neural module networks, as opposed to gradient norms [16] or soft attention maps [27,9].…”

Self-Assembling Modular Networks for Interpretable Multi-Hop Reasoning

“…Neural Modular Network Neural Modular Network (NMN) is a class of models that are composed of a number of sub-modules, where each sub-module is capable of performing a specific subtask. In NMN (Andreas et al, 2016b), N2NMN (Hu et al, 2017), PG+EE (Johnson et al, 2017), GroundNet (Cirik et al, 2018) and TbD (Mascharka et al, 2018), the entire reasoning procedure starts by analyzing the question and decomposing the reasoning procedure into a sequence of sub-tasks, each with a corresponding module. This is done by either a parser (Andreas et al, 2016b;Cirik et al, 2018) or a layout policy (Hu et al, 2017;Johnson et al, 2017;Mascharka et al, 2018) that turns the question into a module layout.…”

Non-monotonic Logical Reasoning and Deep Learning for Explainable Visual Question Answering

“…Object bounding boxes are graph nodes while edges are learned using an attention model conditioned on the question. Mascharka et al [27] augment a deep network architecture with an image-space attention mechanism based on a set of composable visual reasoning primitives that help examine the intermediate outputs of each module. Li et al [23] introduce a captioning model to generate an image's description, reason with the caption and the question to construct an answer, and use the caption to explain the answer.…”