Understanding by Analogy? On the Potential Role of Complexity Science in Explaining Deep Neural Networks

Despite our full access to their internal structure, deep neural networks are epistemically opaque to us; we cannot extract meaningful explanations from existing micro-level descriptions. This paper examines whether complexity science – the systematic use of analogical mappings between models of complex systems – can address this opacity. We reconstruct the methodology of complexity science through the lens of structure-preserving analogies that transfer abstract relational features between source and target domains. Recent Physics of AI approaches that apply effective field theory and renormalization to DNNs exemplify this practice. By analyzing the relationship between analogical model construction and scientific explanation, we argue that complexity science can explain DNNs when analogies guide the development of higher-level models. Thus, our work provides a philosophical basis for evaluating how physical concepts might illuminate artificial intelligence systems. However, due to the heuristic value of such analogies, the resulting models must be validated theoretically and empirically on a case-by-case basis.