Petri nets reveal how binary neural networks make decisions
From Constraints to Causality: Petri net Encodings for Explaining Binary Neural Network Inference
Logic in Computer ScienceMachine Learning
Summary
Binary neural networks are simple AI models that work well on small devices, but they are tricky to understand because their decisions are based on complex, step-by-step calculations. The authors turn these calculations into a form called Petri nets, which show every step in the decision process clearly and in order. This helps explain exactly how changing one input feature can change the decision. Their method can also compare two different neural networks to see where their decisions differ.
What this means in practice
- •For autonomous systems engineers: Trace and explain decisions inside binary neural networks to improve safety and trust in self-driving systems.
- •For embedded hardware developers: Identify precisely where input changes affect binary neural network outputs to optimize on-device AI behavior.
Authors
Alex Chan, Mohamed Tarraf, Alex Yakovlev, Rishad Shafik
Abstract
Binary Neural Networks (BNNs) are attractive for resource-constrained hardware and on-device components in autonomous systems. However, their discrete, non-linear behaviour makes them difficult to analyse and trust in safety-relevant settings. Existing approaches based on SAT and Satisfiability Modulo Theories can verify a BNN's input-output properties, but treat it as a monolithic constraint system, which reveal only whether a property holds and not how the computation reaches its result. In this paper, we present an early work that provides a complementary, discrete-event view. We give a compositional translation from the inference semantics of a fixed, trained BNN to a 1-safe Petri net (PN), where each computation step is an explicit transition, so that a firing sequence is a causal trace of the BNN's decision. On a worked example, we establish faithfulness via exhaustive input-output equivalence against a reference implementation, and show how contrasting firing sequences localises where a single-feature change alters the prediction. Finally, we outline two directions this opens: a correctness-of-translation theorem, and a causal-trace analysis of where two trained BNNs diverge in their decisions, a structural question that constraint-based verifiers cannot express.