Machine learning continues to play a critical role in our society. The ability to automatically identify intricate relationships in large volumes of data has proven incredibly useful for problems such as automatic speech recognition and image processing. In particular, neural networks have become increasingly popular in a wide set of application domains, given their ability to solve complex problems and process high-dimensional data. However, the impressive performance of state-of-the-art neural networks comes at the cost of large area and power consumption for the computation resources used in training and inference. As a result, a growing area of research concerns hardware implementations of neural networks. This work proposes a hardware-friendly design for a time-delay reservoir (TDR), a type of recurrent neural network. TDRs represent one class of reservoir computing neural network topologies, which employ random spatio-temporal feature extraction from time series data in order to produce a linearly separable set of features. Reservoir computing topologies differ from traditional recurrent neural networks because their recurrent weights are fixed, and the only the feedforward output weights need to be trained, usually with linear regression. Previous work on TDRs includes photonic implementation, software implementation, and both digital and analog electronic implementations. This work adds to the body of previous research by exploring the design space of a novel TDR based on single-input signature registers (SISRs), which are common digital circuits used for built-in self-test. The work is motivated by the structural similarity (delayed feedback loop) between TDRs and SISRs, and the possibility of dual-purpose of SISRs for conventional testing as well as machine learning within a single chip. The proposed designs can perform classification on multivariate datasets and perform better than a traditional TDR with quantized reservoir states for parity check, MNIST classification, and temperature prediction tasks. Classification accuracies of up to 100% were observed for some configurations of the SISR for the parity check task and accuracies of up to 85% were observed for MNIST classification. We also observe overfitting on a temperature prediction task with longer data sequences and provide analyses of the results based on the reservoir dynamics, as measured by the rate of divergence between SISR states and the SISR period.

Library of Congress Subject Headings

Machine learning; Neural networks (Computer science); Time delay systems; Registers (Computers)

Publication Date


Document Type


Student Type


Degree Name

Computer Engineering (MS)

Department, Program, or Center

Computer Engineering (KGCOE)


Cory E. Merkel

Advisor/Committee Member

Marcin Lukowiak

Advisor/Committee Member

Dhirseesha Kudithipudi


RIT – Main Campus

Plan Codes