Deep Arguing: A Neurosymbolic Approach for Interpretable Deep Learning

Researchers Adam Gould and Francesca Toni have introduced 'Deep Arguing,' a novel neurosymbolic framework designed to address the lack of interpretability in standard deep learning models. Published on arXiv, this approach integrates deep neural networks with argumentation construction and reasoning to facilitate interpretable classification across various data modalities. Unlike traditional models that tightly couple feature extraction with task objectives, Deep Arguing constructs an argumentation structure where data points actively support their assigned labels while attacking alternative classifications. By employing differentiable argumentation semantics, the model undergoes end-to-end training to jointly learn feature representations and argumentative interactions. This process yields faithful, case-based explanations for predictions, guided by structural constraints on the argumentation graph. Experimental results using both tabular and imaging datasets demonstrate that Deep Arguing achieves predictive performance competitive with standard baselines. Crucially, it offers superior interpretability by providing clear, logical reasoning paths for its decisions, thereby tackling the open challenge of understanding how complex AI models arrive at specific conclusions.