An emulator for biologically-inspired neural networks

This work presents an emulator that has been developed for real-time algorithm and architecture exploration and verification of biologically-inspire neural networks. It can implement a wide range of user-defined neural network types and neuron models. The most complex neuron model is represented by a "biological" neuron that incorporates not only synaptic weighting, postsynaptic summation, static threshold, and saturation, but also other parameters, such as synaptic time delays, neuron gain, computation of membrane potential, and dynamic thresholding, all variable and learnable. For this purpose, a special custom-made CMOS chip has been developed, fabricated, and tested. The chip has been used to build a neural emulator in a form of neural grid array that can interface sensors, actuators, and a host computer.