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

By David Mascharka, Phi Vu Tran, Ryan Soklaski, Arjun Majumdar. 2018

224 citations recorded by the source corpus.

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 state-of-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.

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

Citations in the corpus

Grammar-Based Grounded Lexicon Learning (2022, methodology): “(2) TbD-Net [31] uses a pre-trained semantic parser to parse the question into a symbolic program, and executes the program with a neural module network [5].”
Neural Logic Machines (2019, background): “Besides, modular networks (Andreas et al., 2016; 2017; Mascharka et al., 2018) are proposed for the reasoning over subsymbolic data such as images and natural language question answering.”
A Corpus for Reasoning about Natural Language Grounded in Photographs (2018, background): “Several recent approaches to visual reasoning make use of modular networks (Hu et al., 2017; Johnson et al., 2017b; Mascharka et al., 2018; Perez et al., 2018).”
CLEVRER: CoLlision Events for Video REpresentation and Reasoning (2019, methodology): “TbD-net (Mascharka et al., 2018) follows a similar approach by parsing the input question into a program, which is then assembled into a neural network that acts on the attention map over the image features.”
The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision (2019, methodology): “The SOTA methods do not provide interpretable representation on individual objects (Johnson et al., 2017a; Hudson & Manning, 2018; Mascharka et al., 2018) .”
Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding (2018, background): “These structures include the underlying functional programs [Mascharka et al., 2018, Suarez et al., 2018] and dependency trees [Cao et al., 2018] of the input question.”
MDETR - Modulated Detection for End-to-End Multi-Modal Understanding (2021, background): “The CLEVR dataset consists of 3D-rendered scenes containing between 3 and 10 objects of various shapes, material, size and color.”
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].”
Learning by Abstraction: The Neural State Machine (2019, methodology): “However, in contrast to our approach, those models heavily rely on symbolic execution, either through strong supervision of program annotations [61], non-differentiable python-based functions [87], or a collection of hand-crafted modules speciﬁcally designed for each given task [60], and…”
RAVEN: A Dataset for Relational and Analogical Visual REasoNing (2019, background): “However, superior performance gain was observed in very recent works [6, 36, 58] that fell back to structured representations by using primitives, dependency trees, or logic.”
Answer Them All! Toward Universal Visual Question Answering Models (2019, methodology): “Common approaches include modular networks, with some using ground-truth programs [20, 36], and others learning compositional rules implicitly [17, 18].”
Trends in Integration of Vision and Language Research: A Survey of Tasks, Datasets, and Methods (2019, background): “(Mascharka et al., 2018) proposed a set of visual-reasoning primitives which, when composed, manifest as a model capable of performing complex reasoning tasks in an explicitly interpretable manner.”
Interpretable Visual Question Answering by Reasoning on Dependency Trees (2018, methodology): “Furthermore, PE-Strong [7], NS-VQA [9] and TbD+reg+hres [26] used extra program layouts as additional supervision signals.”
Towards Transparency by Design for Artificial Intelligence (2020, methodology): “Building upon the idea of “Transparency by Design” (TbD) as an emerging concept (Hildebrandt 2013; Mascharka et al. 2018), we aim to provide such a roadmap addressed especially towards those tasked with the development of ADM systems.”
Interpretable Visual Reasoning via Probabilistic Formulation Under Natural Supervision (2020, background): “Consequent studies improve this work by proposing better layout policy [21,20], or designing more speciﬁc modules [42,27].”
MUREL: Multimodal Relational Reasoning for Visual Question Answering (2019, background): “In such methods [17, 14, 29], a neural network reads the question and generates a program, corresponding to a graph of elementary neural operations that process the image.”
Coarse-to-Fine Reasoning for Visual Question Answering (2021, background): “There are numerous reasoning VQA methods [1, 6, 12, 14,19,36,37,39,48,52,55,58,63,64] that focus on learning the relations between visual regions and words in questions implicitly, e.”