An Analysis of Analog Devices and Neural Networks

This thesis investigates analog devices and their implementation in neural networks. It briefly overviews what a neural network is. Then it covers why analog devices are important for neural networks. Several different types of analog devices are shown and described. Then TaOx ReRAM is heavily focused on. Techniques for programming TaOx are described and resulting data is shown. Some models for the different types of errors are utilized. The data and models are combined and implemented into a neural network simulation using these devices. The accuracy for the simulation is shown. Next the analog supporting circuitry is investigated, specifically analog-to-digital converters (ADC). Different device mismatches effects on ADCs and the subsequent effect on neural networks are investigated. Several ADC topologies are presented and the effects of different device mismatch on the ADCs shown. These ADCs are then implemented into a neural network simulator and the accuracy of the network is evaluated. A couple of redesigns are presented that can help mitigate the effects of device mismatch on neural networks. A new metric for evaluating ADC errors that better projects to neural network accuracy is shown.