Date of Award

2011

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Electrical and Computer Engineering

First Advisor

Ahmadi, Majid (Electrical and Computer Engineering)

Keywords

Electrical Engineering.

Rights

CC BY-NC-ND 4.0

Abstract

Artificial neural networks are widely used in many applications such as signal processing, classification, and control. However, the practical implementation of them is challenged by the number of inputs, storing the weights, and realizing the activation function.In this work, Continuous Valued Number System (CVNS) distributed neural networks are proposed which are providing the network with self-scaling property. This property aids the network to cope spontaneously with different number of inputs. The proposed CVNS DNN can change the dynamic range of the activation function spontaneously according to the number of inputs providing a proper functionality for the network.In addition, multi-valued CVNS DRAMs are proposed to store the weights as CVNS digits. These memories scan store up to 16 levels, equal to 4 bits, on each storage cell. In addition, they use error correction codes to detect and correct the error over the stored values.A synapse-neuron module is proposed to decrease the design cost. It contains both synapse and neuron and the relevant components. In these modules, the activation function is realized through analog circuits which are far more compact compared to the digital look-up-tables while quite accurate.Furthermore, the redundancy between CVNS digits together with the distributed structure of the neuron make the proposal stable against process violations and reduce the noise to signal ration.

Share

COinS