Title of Poster / Presentation

Biochemical Reservoir Computing

Location

Portland State University

Start Date

2-5-2018 11:00 AM

End Date

2-5-2018 1:00 PM

Subjects

Neural networks (Computer science) -- Design and construction, Knowledge representation (Information theory), Natural computation

Abstract

Reservoir computing is an emerging machine learning paradigm. Compared to traditional feedforward neural networks, the reservoir can be unstructured and recurrent and only the output layer is trained. Reservoirs can be built with various types of physical components, yet, biochemical building blocks have not been widely used. This project focuses on designing and testing a reservoir computer (RC) based on chemical reaction network (CRN). We simulated high-level CRNs in MATLAB and their complex chemical dynamics were observed over time. A CRN constructed by a network of coupled deoxyribozyme oscillators was chosen for the final RC model. The inputs of the RC were perturbed by the influx rates of the substrate molecules. With the perturbations, this reservoir showed dynamics of an amplifying and attenuating oscillator. Random perturbations were introduced at different times to produce random oscillations of the product concentrations. A readout layer of the RC was implemented with a single perceptron that was trained to learn Boolean functions and the Hamming distance between two input bitstreams. From the experiments, the results show that such an RC can learn linearly separable patterns, with an average learning error of 0.0031. For future work, this biochemical RC can be transformed into DNA strands in order to detect pathogens or gene mutations. This project thus has the potential to influence biomedical research and genetic disorder treatments.

Persistent Identifier

http://archives.pdx.edu/ds/psu/25055

Share

COinS
 
May 2nd, 11:00 AM May 2nd, 1:00 PM

Biochemical Reservoir Computing

Portland State University

Reservoir computing is an emerging machine learning paradigm. Compared to traditional feedforward neural networks, the reservoir can be unstructured and recurrent and only the output layer is trained. Reservoirs can be built with various types of physical components, yet, biochemical building blocks have not been widely used. This project focuses on designing and testing a reservoir computer (RC) based on chemical reaction network (CRN). We simulated high-level CRNs in MATLAB and their complex chemical dynamics were observed over time. A CRN constructed by a network of coupled deoxyribozyme oscillators was chosen for the final RC model. The inputs of the RC were perturbed by the influx rates of the substrate molecules. With the perturbations, this reservoir showed dynamics of an amplifying and attenuating oscillator. Random perturbations were introduced at different times to produce random oscillations of the product concentrations. A readout layer of the RC was implemented with a single perceptron that was trained to learn Boolean functions and the Hamming distance between two input bitstreams. From the experiments, the results show that such an RC can learn linearly separable patterns, with an average learning error of 0.0031. For future work, this biochemical RC can be transformed into DNA strands in order to detect pathogens or gene mutations. This project thus has the potential to influence biomedical research and genetic disorder treatments.