Explainable Neural Computation via Stack Neural Module Networks

By Ronghang Hu, Jacob Andreas, Trevor Darrell, Kate Saenko. 2018

210 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Model Architectures, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Model Architectures

In Safety, Robustness & Interpretability

Citations in the corpus

Visual Programming: Compositional visual reasoning without training (2022, methodology): “Neural module networks (NMN) [2] pioneered modular and compositional approaches for the visual question answering (VQA) task.”
From Recognition to Cognition: Visual Commonsense Reasoning (2018, background): “This has motivated work in having models provide explanations for their behavior, in the form of a natural language sentence [31, 9, 41] or an attention map [32, 35, 37].”
A Corpus for Reasoning about Natural Language Grounded in Photographs (2018, background): “…including modular neural networks (Andreas et al., 2016a,b; Johnson et al., 2017b; Perez et al., 2018; Hu et al., 2017; Mascharka et al., 2018; Hu et al., 2018; Suarez et al., 2018) and attention-or memory-based meth-ods (Santoro et al., 2017; Zhu et al., 2017; Hud-son and Manning, 2018; Tan…”
Learning Visual Question Answering by Bootstrapping Hard Attention (2018, methodology): “However, only soft attention is used in the majority of Visual QA works [7,8,9,10,11,12,43,44,45,46,47,48,49,50,51,52].”
Meta Module Network for Compositional Visual Reasoning (2019, methodology): “Stack-NMN [14] proposes to make soft layout selection so that the whole model is fully differentiable, and Neural-Symbolic VQA [44] proposes to perform completely symbolic reasoning by encoding images into scene graphs.”
Separating Skills and Concepts for Novel Visual Question Answering (2021, unknown): “Other efforts create compositional models to handle the relational reasoning chains [23, 24, 26, 28, 50].”
A Competence-aware Curriculum for Visual Concepts Learning via Question Answering (2020, methodology): “ the supervision or increasing the generalization ability to new tasks in NMNs. For example, Johnson et al. [31] replaces the hand-designed syntactic parsers by a learned program generator. Stack-NMN [24] can be trained without module layout supervision using a stack-based data structure. Neural-Sym”
Grounding Physical Concepts of Objects and Events Through Dynamic Visual Reasoning (2021, methodology): “To perform deeper visual reasoning, neural module networks were extensively studied in recent works (Johnson et al., 2017a; Hu et al., 2018; Hudson & Manning, 2018; Amizadeh et al., 2020), where they represent symbolic operations with small neural networks and perform multi-hop reasoning.”
Obtaining Faithful Interpretations from Compositional Neural Networks (2020, background): “Work on NMNs thus far (Hu et al., 2018; Jiang and Bansal, 2019) has overlooked systematically evaluating faithfulness, performing only qualitative analysis of intermediate outputs.”
Trends in Integration of Vision and Language Research: A Survey of Tasks, Datasets, and Methods (2019, background): “Another approach modifies Neural modular networks (Hu et al., 2018) such that it performs compositional reasoning by automatically inducing a desired sub-task decomposition without relying on strong supervision.”
Learning to Collocate Neural Modules for Image Captioning (2019, background): “However, we believe that a combination of simple neural modules can approximate to accomplish a variety of complex tasks [19].”
Dynamic Visual Reasoning by Learning Differentiable Physics Models from Video and Language (2021, background): “Existing works can be generally categorized into two streams as end-to-end approaches [38, 81, 72, 39, 5] and neurosymbolic approaches [78, 27, 28, 22, 4, 57, 43, 56, 34, 3].”
Learning to Assemble Neural Module Tree Networks for Visual Grounding (2018, background): “Although some of the performance gains are marginal, one should notice that it seems NMTREE balances the well-known trade-off between performance and explainability [10].”
RUBi: Reducing Unimodal Biases in Visual Question Answering (2019, background): “Current end-to-end VQA models [15, 16, 17, 18, 19, 20, 21, 22] achieve impressive ⇤Equal contribution”
Self-Assembling Modular Networks for Interpretable Multi-Hop Reasoning (2019, methodology): “To make the entire model differentiable, we also adopt soft program execution from Hu et al. (2018), where the output of each step is the sum of outputs from all modules weighted by the module distribution (blue-shaded area in Fig.”
Dense Regression Network for Video Grounding (2020, background): “Since the model may focus on different parts of the input query at each level, we follow [18] to fuse the query and the video features at different levels.”
Chameleon: Plug-and-Play Compositional Reasoning with Large Language Models (2023, background): “Some later work [16, 13, 12] takes a step further by predicting instance-specific network layouts in an end-to-end manner, without relying on parsers, using reinforcement learning [50] and weak supervised learning.”