Interpreting Neural Networks as Quantitative Argumentation Frameworks

Nico PotykaAAAI 2021 · DOI

Abstract

This paper establishes a formal mathematical bridge between Multi-Layer Perceptrons (MLPs) and Quantitative Bipolar Argumentation Frameworks (QBAFs). By demonstrating that neural networks can be interpreted as systems of "pro" and "contra" arguments, the research provides a novel lens for deep learning interpretability and a robust framework for learning argumentation parameters via standard backpropagation.

Introduction

Neural networks, particularly Multi-Layer Perceptrons (MLPs), are renowned for their predictive power but often criticized for their "black box" nature. Conversely, Formal Argumentation offers transparent, symbolic reasoning through graphs of "attacks" and "supports," yet these models are typically difficult to train on raw data. This paper addresses this gap by asking whether the mathematical operations of a neural network can be interpreted as the process of weighing arguments and whether the rigorous "desirable properties" of argumentation can improve our understanding of neural models.

The Mapping: MLPs as Argumentation Frameworks

The core contribution is the introduction of MLP-based semantics for Quantitative Bipolar Argumentation Frameworks (QBAFs). In this mapping, an MLP is viewed as a QBAF where the "aggregation" of arguments corresponds to the weighted sum of inputs, and the "influence" function—determining how an argument's strength changes—corresponds to the neural network's activation function (specifically the logistic/sigmoid function).

Architecture of a Multi-Layer Perceptron (MLP) and the mathematical operation of a single neuron.

Under this framework, a hidden layer in an MLP is not merely a collection of abstract features; it represents a set of intermediate arguments being supported or attacked by the layer below it. This allows for a symbolic interpretation of internal weights.

Handling Cyclic Graphs and Continuous Semantics

While standard feed-forward networks are acyclic, the author extends this interpretation to cyclic graphs where arguments can form feedback loops. To resolve convergence issues—where iterative updates might oscillate indefinitely—the paper proposes a continuous variant using a system of differential equations. This ensures that argument strengths stabilize even in complex, loopy structures. A key theoretical result is the "Convergence Bound": if the product of the maximum edge weight ($W$) and the maximum number of parents ($P$) satisfies $W \cdot P < 4$, the system is guaranteed to converge to a single stable interpretation.

A Quantitative Bipolar Argumentation Framework (QBAF) representing a Buy/Sell decision alongside its formal semantics.

Evaluation and Key Results

The author evaluated the MLP-based semantics against 12 "desirable properties" established in argumentation literature, such as Neutrality, Monotony, and Reinforcement. Surprisingly, the MLP-based semantics satisfied these formal properties better than many semantics specifically designed for argumentation, such as Euler-based methods. Notably, it satisfies "Duality," treating attacks and supports symmetrically.

Furthermore, because the mapping is differentiable, it enables end-to-end learning. Standard deep learning libraries can be used to "learn" the base scores and weights of an argumentation framework directly from data, effectively combining human expertise (in the form of graph structure) with data-driven optimization.

Conclusion and Future Work

While this mapping provides a path toward interpretable MLPs, dense networks remain challenging for human readability. The most promising application lies in sparse networks, where an expert defines an initial argumentation graph that is subsequently fine-tuned via backpropagation. Future work will focus on these "Sparse MLPs" to bridge the gap between expert knowledge and empirical performance. One minor limitation is "closed-mindedness" at the boundaries (base scores of 0 or 1), which can be mitigated by using values like 0.01 and 0.99 in practice.

Summary updated

BibTeX
@inproceedings{potyka2021interpreting,
  author = {Nico Potyka},
  title = {{Interpreting Neural Networks as Quantitative Argumentation Frameworks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i7.16801},
}
Figure