Article-Journal

It is unclear if biological qualities, or constraints, impact a learning and computing neuron model positively, negatively, or at all. We simulate models with and without these constraints. We find models with a combination of biological constraints can outperform models with isolated introduced constraints. Qualitative dendritic properties may contribute inductive biases to neuron models that learn and compute.

It is unclear how qualitative dendritic properties contribute to neural computation beyond nonlinearity. We investigate the impact of thee properties in a neuron model derived from an artificial neural network. We find that these properties can negatively and positively impact performance on benchmark computer vision tasks. This opens questions about the potentially beneficial inductive biases that dendritic properties might introduce to a learning and computing model. Previously titled /“Can a Single Neuron Solve MNIST?/”