KIP-Veröffentlichungen

To achieve their goal of realizing fast and energy-efficient learning, neuromorphic systems require computationally powerful models that obey the constraints imposed by a physical implementation of neural network structure and dynamics, such as the inevitability of relaxation times or the locality of plasticity. In this work, we provide a first-principles derivation of a mechanistic model for cortical computation based on the premise of "neuronal least action".  The resulting time-continuous neuron and synapse dynamics realize gradient-descent learning through error backpropagation both in supervised and in reinforcement learning scenarios. In particular, the derived equations of motion reproduce well-established microscopic phenomena such as neuronal leaky integration of afferent signals, while enabling synaptic learning using only locally available information. Our principled framework can thus serve as a starting point for hardware-focused models of highly efficient time-continuous learning.