Going deeper with morphologically detailed neural networks by simulation-based gradient propagation

Morphologically detailed dendrites possess powerful computational capabilities but are computationally expensive. Therefore, exploring how large-scale, detailed multi-compartment neural networks achieve learning proves challenging, presenting significant obstacles both in learning algorithms and simulation infrastructures. Here, we provide an extension to the DeepDendrite framework to enable the construction and data-driven training of multi-layer, detailed multi-compartment neural networks with highly modularized layer components. The gradient at each layer is computed by simulating gradient-mirror neurons in the feedback pathway simultaneously with the detailed neurons in the feedforward pathway, providing a simulation-based implementation of backpropagation in detailed neural networks. We demonstrate comparable results on classic image classification datasets with fully-connected and convolutional architectures. Furthermore, we analyze transfer attack robustness between artificial neural networks (ANNs), detailed neural networks and single-compartment networks. In conclusion, we provide a useful framework to investigate learning in multi-layer detailed neural networks, potentially offering further insights into exploiting the computational potential of dendrites at a large scale.