Analog Neural Network Integrated Circuit

Software models of neural networks (NN) these days are used everywhere in artificial intelligence (AI) applications. As they are computationally rather resource and power demanding, hardware support for neural networks is in high demand. GPUs (i.e. graphic cards) for example, have an architecture that can be exploited for efficient neural network computations. THIS project however, wants to be even more power efficient, and implement neural networks as analog integrated circuits.
A key feature of NNs are the synaptic weights, the parameters that decide the NN's behaviour and that are modified by learning. An analog NN implementation thus needs a means of storing analog weights. In the past, our group has used Flash memory not for digital but for analog non-volatile storage, where non-volatile means that the content of that memory is preserved also when the power supply is turned off. Flash memory CAN be used for analog storage since it is basically charge stored on a 'floating gate', and this charge can be controlled very precisely also for analog storage.

This project shall investigate the use of flash memory for analog storage in a specific integrated circuit CMOS technology and look into building a simple NN structure using this analog memory. It is part of a bigger project for neuromorphic (computing inspired by the nervous system) sound recognition, where others concentrate on feature extraction of input sound to produce an input feature vector for the NN and on the learning algorithm to learn the appropriate NN synaptic weight.