Training of artificial neural networks

Apparatus for implementing an artificial neural network, having a succession of neuron layers with interposed synaptic layers each storing a respective set of weights {w} for weighting signals propagated between its adjacent neuron layers, in an iterative training cycle of signal propagation and weight-update calculation operations, the apparatus comprising:

• a processor adapted to generate signals propagated by the neuron layers in said propagation operations and to perform said weight-update calculation operations;
• and
• a memcomputing unit storing said weights in respective memristive devices for receiving and weighting signals propagated by the neuron layers to implement the synaptic layers of the artificial neural network, the memcomputing unit including a programming circuit for programming the memristive devices and, for at least one of the synaptic layers, a plurality of P1 arrays of memristive devices, each array storing the set of weights of that synaptic layer Sl in respective memristive devices;
• wherein the memcomputing unit is adapted, in said signal propagation operation, to supply respective subsets of the signals to be weighted by the synaptic layer Sl in parallel to the P1 arrays;
• wherein the processor is adapted, in said weight-update calculation operation, to calculate updates to respective weights stored in each of the P1 arrays in dependence on signals propagated by the neuron layers, and, for each weight w in said set, to accumulate the updates calculated for that weight in the P1 arrays in an accumulation value xw for that weight;
• wherein the processor is adapted to control said programming circuit to periodically program the memristive devices storing each weight w in all of the P1 arrays to update a stored weight in dependence on the accumulation value xw for that weight;
• and wherein the apparatus is adapted:
• to program the memristive devices storing said weight w in the P1 arrays by applying a number q, dependent on said accumulation value xw, of programming pulses to the memristive devices;
• and
• to update a magnitude of the accumulation value Xw for that respective weight w to JXwl−qE, where F is a predetermined step-size indicative of weight-change of a memristive device in response to application of said programming pulse.