Artificial Neural Network (ANN), a computational model based on the biological neural networks, has a recent resurgence in machine intelligence with breakthrough results in pattern recognition, speech recognition, and mapping. This has led to a growing interest in designing dedicated hardware substrates for ANNs with a goal of achieving energy efficiency, high network connectivity and better computational capabilities that are typically not optimized in software ANN stack. Using stochastic computing is a natural choice to reduce the total system energy, where a signal is expressed through the statistical distribution of the logical values as a random bit stream. Generally, the accuracy of these systems is correlated with the stochastic bit stream length and requires long compute times.

In this work, a framework is proposed to accelerate the long compute times in stochastic ANNs. A GPU acceleration framework has been developed to validate two random projection networks to test the efficacy of these networks prior to custom hardware design. The networks are stochastic extreme learning machine, a supervised feed-forward neural network and stochastic echo state network, a recurrent neural network with online learning. The framework also provisions identifying optimal values for various network parameters like learning rate, number of hidden layers and stochastic number length. The proposed stochastic extreme learning machine design

is validated for two standardized datasets, MNIST dataset and orthopedic dataset. The proposed stochastic echo state network is validated on the time series EEG dataset. The CPU models were developed for each of these networks to calculate the relative performance boost. The design knobs for performance boost include stochastic bit stream generation, activation function, reservoir layer and training unit of the networks. Proposed stochastic extreme learning machine and stochastic echo state network achieved a performance boost of 60.61x for Orthopedic dataset and 42.03x for EEG dataset with 2^12 bit stream length when tested on an Nvidia GeForce1050 Ti.

Library of Congress Subject Headings

Neural networks (Computer science); Graphics processing units; Machine learning; Learning models (Stochastic processes)

Publication Date


Document Type


Student Type


Degree Name

Computer Engineering (MS)

Department, Program, or Center

Computer Engineering (KGCOE)


Dhireesha Kudithipudi

Advisor/Committee Member

Marcin Lukowiak

Advisor/Committee Member

Sonia Lopez Alarcon


RIT – Main Campus

Plan Codes