Research

Thesis Overview

Computers were an imitation of human brain from the very beginning. Though we were aware of the neural architecture of the brain, technology at the time favored the von-Neumann architecture. While the memory and processing are at the same physical location in a biological neural network, they are separate in the von-Neumann architecture. This meant constant data shuffling and were a major source of energy inefficiency. Supercomputers consuming millions of Watts still could not perform tasks that human brain can with mere 20 Watts.

Simply adopting the neural architecture was not sufficient. The energy efficiency problem demanded, we also address how we implement them on the physical hardware. This is where Neuromorphic engineering and In-Memory Computing came in. Both of these approaches attempts to process signals uing the physical characteristics of arrays of memory devices.

In this thesis, we brings to light the key challenges faced in this field, and directly addresses the precision limits of these memory device by experimental demonstration of several architecture modifications to maintain high accuracy.