Source-linked AI summary
Explainable Neural Computation via Stack Neural Module Networks
Ronghang Hu, Jacob Andreas, Trevor Darrell, Kate Saenko
TL;DR
Complex reasoning models need compositional procedures whose intermediate decisions are interpretable, but prior modular approaches commonly rely on expert layout supervision. The paper introduces Stack-NMN, which induces soft module layouts and shares modules across VQA and REF without that supervision. Experiments report comparable modular performance and stronger human understanding of the model’s reasoning and likely successes or failures.
Problem
Prior modular reasoning models commonly require expert layouts, while the interpretability of complex multi-step reasoning remains insufficiently evaluated with human studies.
Method
Stack-NMN uses soft, continuous module layouts with differentiable training and shared neural modules for Visual Question Answering and Referential Expression Grounding.
Results
Stack-NMN has comparable performance to existing modular approaches without expert supervision and achieves the greatest interpretability among evaluated models on subjective and objective human-understanding measures.
Takeaways & Limitations
Intermediate module choices, image attentions, and textual attentions provide reasoning explanations that help users understand the model and predict whether it will succeed or fail.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper targets complex visual reasoning that requires multiple subtasks and interpretable intermediate decisions. Stack-NMN induces soft modular reasoning without expert layout supervision and supports both VQA and REF.
- Motivation: Complex visual questions can require finding referenced objects, reasoning about spatial relations, and deciding whether the queried object exists.A single spatial heat map may not reveal this multi-step reasoning procedure.
- Approach: Stack-NMN uses an explicit modular reasoning process trained end-to-end with back-propagation and without expert supervision of reasoning steps.It replaces discrete layout choices with soft, continuous module weights.
- Approach: The model shares neural modules across Visual Question Answering and Referential Expression Grounding, linking related tasks through common sub-routines.This enables both tasks to be handled within a single model.
- Evaluation: The evaluation asks whether explicit modular structure improves interpretability as well as accuracy on complex reasoning tasks.CLEVR is used as a testbed, alongside comparisons with models such as Relation Networks, FiLM, and MAC.
- Findings: Stack-NMN achieves the greatest interpretability among evaluated models on subjective and objective measures of human understanding.The reported comparison covers human understanding of the model’s reasoning process and predictions.
2 Related work
Prior work spans holistic and modular approaches to multi-step reasoning, but modular systems commonly depend on expert layouts. This paper instead uses shared modules across related tasks and soft continuous layouts to avoid that requirement while evaluating interpretability directly.
- Existing approaches: Holistic models lack explicit sub-task structure, whereas neural module networks decompose reasoning into sub-tasks handled by specialized modules.Sequential attention models can perform multiple reasoning steps without exposing the same structured module decomposition.
- Existing approaches: Earlier neural module networks use layout policies to convert questions into module sequences that are then executed by neural modules.The layout policy determines which sub-tasks to perform, while modules learn how to perform them.
- Open problem: Previous modular systems often require human-annotated expert layouts because learning what to do and how to do it simultaneously is difficult.Without expert guidance, they can suffer significant performance drops or fail to converge.
- Interpretability: The paper evaluates whether explicit modular structure improves interpretability in concrete human-evaluation scenarios rather than assuming modularity is interpretable.The comparison focuses on users’ understanding of internal reasoning and predictions.
- Multi-task learning: The model handles VQA and REF by sharing a common set of neural modules for related sub-routines rather than only sharing common features.This frames multi-task learning around shared reasoning routines.
3 Approach
Stack-NMN uses a differentiable controller to induce a soft module layout, execute neural modules over a memory stack, and support compositional reasoning without discrete layout selection.
- 3 Approach: The model combines a layout controller, differentiable neural modules, and a memory stack for modular reasoning.The controller specifies subtasks, modules perform them, and the stack stores intermediate outputs.
- 3.1 Module layout controller: At each time step, the controller selects modules with continuous weights and supplies textual parameters that specify the subtask instruction.The module weights form a soft distribution over modules, while textual attention produces the parameter c_t.
- 3.1 Module layout controller: The controller encodes the question with a bidirectional LSTM and recurrently updates its state to predict module weights and textual parameters.A time-dependent transform combines the question representation with the previous textual parameter before module selection.
- 3.2 Neural modules with a memory stack: Neural modules are differentiable functions for specific subtasks, including layouts that compose Find, Transform, and Answer operations.The implementation merges several answering and comparison modules and adds NoOp for padding layouts to a maximum length.
- 3.2 Neural modules with a memory stack: A differentiable LIFO memory stack stores intermediate image-attention outputs so modules can pop prior results and push updated results in tree-structured layouts.The stack uses a memory array and pointer with differentiable push and pop operations.
- 3.3 Soft program execution: At each step, the model executes every module, averages their resulting memory stacks according to the predicted weights, and sharpens the stack pointer.This weighted execution implements continuous rather than discrete layout selection.
4 Experiments
Experiments evaluate Stack-NMN on VQA and REF, testing task performance, layout induction without expert supervision, shared sub-tasks, and human interpretability. The model performs strongly without expert layouts, induces near-discrete module choices, and is more interpretable to humans, although it underperforms MAC on VQA accuracy.
- Experimental setup: Stack-NMN is evaluated on CLEVR VQA and a newly constructed CLEVR-Ref dataset for referential expression grounding.CLEVR-Ref matches CLEVR's scale and requires localizing the referred object with a bounding box.
- Model performance: The model achieves over 90% accuracy on both tasks when trained separately, with or without expert layout supervision.Joint training can produce even higher accuracy, especially without expert layout supervision.
- Sub-task decomposition and layout induction: Stack-NMN induces reasonable sub-task decompositions without expert layouts and shares routines such as find and transform across VQA and REF.The visualized layouts, intermediate outputs, and attended words provide evidence of this cross-task decomposition.
- Sub-task decomposition and layout induction: After convergence, module weights become highly concentrated: over 98% without layout supervision and over 99.99% with supervision.The corresponding average entropies are 0.082 and 7.5 × 10^-5, respectively.
- Model performance: Without expert layout supervision, Stack-NMN significantly outperforms N2NMN, while PG+EE and TbD fail to converge or cannot be trained.The authors attribute this advantage to Stack-NMN's fully differentiable soft layout and back-propagation training.
- Evaluation of interpretability: Human evaluators find Stack-NMN more interpretable and can predict its success or failure above chance, but it underperforms MAC in VQA accuracy.With test-time layout discretization, CLEVR validation accuracy drops by less than 4%.
5 Conclusion
The paper proposes a model for visual question answering and referential expression grounding that shares neural modules across related tasks. It induces inference sub-task decompositions without expert layout supervision and produces explanations that improve human understanding in subjective and objective evaluations.
- The model addresses visual question answering and referential expression grounding by sharing neural modules across related sub-tasks.
- Stack-NMN induces an inference decomposition into sub-tasks without requiring expert layout supervision.
- The model explains reasoning through soft module choices, image attentions, and textual attentions.
- These explanations produced better human understanding in subjective and objective evaluations, despite no human-provided explanations during training.
Explainable Neural Computation via Stack Neural
The supplementary material includes a section titled “Module Networks.”
- The supplementary material contains a section titled “Module Networks.”
- “Module Networks” is identified as supplementary material.
- The supplied passage provides the section title but no further content.
A Visualization of our model and MAC
The visualization section illustrates the model’s differentiable memory stack and compares reasoning-step visualizations with MAC. It includes image and textual attentions alongside selected modules for the proposed model.
- The differentiable memory stack is illustrated through a visualized example in Figure 4.
- The visualization compares reasoning steps from the proposed model with those from the MAC model.
B Details of the human evaluation on interpretability
The human evaluation used Amazon Mechanical Turk to assess subjective understanding and forward prediction of model success or failure. Evaluators viewed visualized reasoning examples from the proposed model and MAC.
- The human evaluation was deployed on Amazon Mechanical Turk.
- For subjective understanding, evaluators judged whether each model’s reasoning steps were clear and understandable.
- The subjective-understanding evaluation showed 200 visualized examples of each model.
- The forward-prediction evaluation used successful and failure examples from each model to assess failure detection.
- The visualized comparisons include models trained with and without expert layout supervision, alongside MAC with 12 steps.
- Figures 6 and 7 show the evaluation interfaces for subjective understanding of the proposed model and MAC.
- Figures 8 and 9 show the evaluation interfaces for forward prediction on the proposed model and MAC.