Global explanations of Bayesian neural networks through local contributions

Eirik Høyheim, Norwegian Defence Research Establishment (FFI)

Abstract: Neural networks are often considered non-interpretable due to their numerous weights and non-linear activation functions. Additionally, their structures are typically chosen arbitrarily through empirical experiments, risking unnecessary complexity. This paper demonstrates that latent binary Bayesian neural networks can drastically reduce the size and complexity while maintaining a well-performing model. We also show that using piecewise linear activation functions allows for the retrieval of interpretable slope coefficients, which can provide both global and local explanations for model predictions.