Interpretability in reinforcement learning is crucial for ensuring AI systems align with human values and fulfill the diverse related requirements including safety, robustness and fairness. Building on recent approaches to encouraging sparsity and locality in neural networks, we demonstrate how the penalisation of non-local weights leads to the emergence of functionally independent modules in the policy network of a reinforcement learning agent. To illustrate this, we demonstrate the emergence of two parallel modules for assessment of movement along the X and Y axes in a stochastic Minigrid environment. Through the novel application of community detection algorithms, we show how these modules can be automatically identified and their functional roles verified through direct intervention on the network weights prior to inference. This establishes a scalable framework for reinforcement learning interpretability through functional modularity, addressing challenges regarding the trade-off between completeness and cognitive tractability of reinforcement learning explanations.

Reservoir computing is a machine learning paradigm that transforms the transient dynamics of high-dimensional nonlinear systems for processing time-series data. Although reservoir computing was initially proposed to model information processing in the mammalian cortex, it remains unclear how the non-random network architecture, such as the modular architecture, in the cortex integrates with the biophysics of living neurons to characterize the function of biological neuronal networks (BNNs). Here, we used optogenetics and fluorescent calcium imaging to record the multicellular responses of cultured BNNs and employed the reservoir computing framework to decode their computational capabilities. Micropatterned substrates were used to embed the modular architecture in the BNNs. We first show that modular BNNs can be used to classify static input patterns with a linear decoder and that the modularity of the BNNs positively correlates with the classification accuracy. We then used a timer task to verify that BNNs possess a short-term memory of ~1 s and finally show that this property can be exploited for spoken digit classification. Interestingly, BNN-based reservoirs allow transfer learning, wherein a network trained on one dataset can be used to classify separate datasets of the same category. Such classification was not possible when the input patterns were directly decoded by a linear decoder, suggesting that BNNs act as a generalization filter to improve reservoir computing performance. Our findings pave the way toward a mechanistic understanding of information processing within BNNs and, simultaneously, build future expectations toward the realization of physical reservoir computing systems based on BNNs.

Modularity of neural networks -- both biological and artificial -- can be thought of either structurally or functionally, and the relationship between these is an open question. We show that enforcing structural modularity via sparse connectivity between two dense sub-networks which need to communicate to solve the task leads to functional specialization of the sub-networks, but only at extreme levels of sparsity. With even a moderate number of interconnections, the sub-networks become functionally entangled. Defining functional specialization is in itself a challenging problem without a universally agreed solution. To address this, we designed three different measures of specialization (based on weight masks, retraining and correlation) and found them to qualitatively agree. Our results have implications in both neuroscience and machine learning. For neuroscience, it shows that we cannot conclude that there is functional modularity simply by observing moderate levels of structural modularity: knowing the brain's connectome is not sufficient for understanding how it breaks down into functional modules. For machine learning, using structure to promote functional modularity -- which may be important for robustness and generalization -- may require extremely narrow bottlenecks between modules.