First Advisor

Christof Teuscher

Term of Graduation

Winter 2020

Date of Publication

3-3-2020

Document Type

Thesis

Degree Name

Master of Science (M.S.) in Electrical and Computer Engineering

Department

Electrical and Computer Engineering

Language

English

Subjects

Neural networks (Computer science), Bioinformatics

DOI

10.15760/etd.7376

Physical Description

1 online resource (xv, 101 pages)

Abstract

Increasing viral illnesses threatens global human health and welfare. Due to the distribution of disease and the expense of diagnosis, it is of value to develop portable assays that can detect viral infections early. DNA molecular logic technology offers a portable detection method due to the versatility and stability of DNA and the potential of in situ computation.

Top-down engineering of these chemical logic networks can be difficult due to the difficulties of their implementation using DNA as a substrate. In this work echo state networks, a form of recurrent neural networks, were explored with the motivation that their implementation in situ would be more feasible. Echo state networks consist of a fixed recurrent hidden layer, called the reservoir, with a trainable linear readout layer. We explored the size and sparsity of these networks with an aim for minimum complexity, small size and fewer connections, while still achieving improvements over linear baseline models.

The performance of these networks was compared against the NARMA benchmark time series prediction task and the classification tasks of digit recognition, gene splice junctions, and dengue virus genome serotypes. The networks were able to perform the tasks for NARMA achieving improvements over memoryless linear models with only 2 nodes and over linear models with input memory with 100 nodes. For the digit recognition task, a 20 node and 10% connectivity network improved performance over linear models. For the genome sequence classification tasks these networks could achieve >90% accuracy on each providing improvements from linear models of 2.8% for the splice junctions and 13.6% for dengue virus classification.

To bridge the model of echo state networks to chemical reaction networks, an oscillating recurrent neural network was simulated which was based on the graph of an oscillating chemical reaction network. Performance was characterized across size and spectral radius of the network. For the NARMA task, an improvement in mean squared error of near an order of magnitude could be achieved by increasing the oscillator from having 6 nodes to having 300 nodes and a similar improvement in performance between a spectral radius of 0.1 and 0.9 was observed. This was also compared to a simulated chemical oscillator in terms of prediction accuracy and classification on the previous tasks.

The simulated results from the classifier suggest that a chemical reaction network of somewhat unknown connectivity and limited complexity but sufficient dynamics could be a potentially easier means of implementing an in situ viral classifier than the traditional methods of top-down engineering. This method will be transferable to other pathogens, both viral and bacterial, using available genome databases.

Rights

In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).

Persistent Identifier

https://archives.pdx.edu/ds/psu/33487

Download

Share

COinS