Presenter Information

Nithyakalyani SampathFollow

Presentation Type

Poster

Start Date

4-5-2022 11:00 AM

End Date

4-5-2022 1:00 PM

Subjects

Memristors, Memcapacitors, Neuromorphic computing

Advisor

Dr. Christof Teuscher

Student Level

Masters

Abstract

Data-intensive computing operations, such as training neural networks, are essential but energy-intensive. Memcapacitance and memristance,which can be described as capacitance and resistance, with “memory”, are properties of semiconductor devices that are observed on the nano-scale. These properties allow for data storage without a constant source of power, leading to hardware which is more energy efficient.

We intend to demonstrate that we can build specialized hardware onto which a neural network can be directly mapped using memristors and memcapacitors, improving the energy efficiency of the network. We will use Simulation Program with Integrated Circuit Emphasis (SPICE) to model our memcapacitor and memristor. Using this model, we will create a reservoir of memristive and memcapacitive elements and evaluate our design across a range of memcapacitors to memristor ratios, while testing reservoir structures including small-world, crossbar, random, hierarchical, and power-law implementations. We hypothesize that our design will greatly improve the energy efficiency and performance of neural networks.

Persistent Identifier

https://archives.pdx.edu/ds/psu/37476

Share

COinS
 
May 4th, 11:00 AM May 4th, 1:00 PM

Memristors, Memcapacitors and Their Application in Neuromorphic Computing

Data-intensive computing operations, such as training neural networks, are essential but energy-intensive. Memcapacitance and memristance,which can be described as capacitance and resistance, with “memory”, are properties of semiconductor devices that are observed on the nano-scale. These properties allow for data storage without a constant source of power, leading to hardware which is more energy efficient.

We intend to demonstrate that we can build specialized hardware onto which a neural network can be directly mapped using memristors and memcapacitors, improving the energy efficiency of the network. We will use Simulation Program with Integrated Circuit Emphasis (SPICE) to model our memcapacitor and memristor. Using this model, we will create a reservoir of memristive and memcapacitive elements and evaluate our design across a range of memcapacitors to memristor ratios, while testing reservoir structures including small-world, crossbar, random, hierarchical, and power-law implementations. We hypothesize that our design will greatly improve the energy efficiency and performance of neural networks.