Evolving interpretable neural modularity in free-form multilayer perceptrons through connection costs.

Interpretability fosters trust when humans and artificial intelligence (AI) systems interact, and its value for neural networks in particular cannot be overstated. In this paper, we analyse the emergence and value of neural modularity as it relates to interpretability. We compare the modularity evolved through connectivity constraints in terms of network Q-scores and examine the interpretable qualities of the resultant networks with functional subset regression. The connectivity constraints compared here include those proposed by previous research as well as several novel variations formulated to express neuron input competition and connections per neuron variance. Networks were evolved using HyperNEAT on a free-form substrate. The results indicate that the connection costs successfully promote the evolution of neural modularity across a variety of tasks and show that the novel connection cost variations are competitive with previously explored connection costs. The interpretability assessment shows that while the evolved networks' interpretable qualities are task dependent, two of the compared connection costs deliver statistically different functional module overlap distributions. However, recovered subnetwork module accuracies remain low, highlighting the key points for future research. [ABSTRACT FROM AUTHOR]