Master thesis

back to list
Natural gradient descent for spiking neuromorphic systems
In complex energy landscapes, naive (Euclidean) gradient descent is often not the most efficient way to reach the (local) energy minimum. The natural gradient represents an efficient and parametrization-invariant solution to efficient learning and can be formulated in a way that is compatible with biological neuro-synaptic dynamics. Due to its invariance properties, it might be particularly useful for substrates that exhibit component diversity, such as analog neuromorphic circuits. This Masterís project is aimed at studying the behavior of spiking natural gradient descent under fixed-pattern noise and its demonstration on one of the neuromorphic platforms developed in the Electronic Vision(s) group.
more information