Machine Learning in Engineering Applications
Room and Cryogenic Temperature Circuit Optimization for Neuromorphic Circuit:
Machine learning (ML), in particular deep-learning, has become an indispensable part of our daily life. Machine learning requires many computing resources and is typically performed on large servers. For edge devices, such as hand-held devices and Internet-of-Thing (IoT), analog circuits utilizing novel electronics, such as Resistive Memory (ReRAM), are needed for edge ML. Such circuits that mimic the neuron behavior and have ultra-low power consumption, are called neuromorphic circuits. Edge ML is also inevitable when data privacy is important. ReRAM neuromorphic circuits are still in their novel stage and a co-optimization of the ReRAM device and other analog circuit components is required to achieve the best trade-off in performance, power, area, and cost (PPAC).
Objectives:
One objective is to study the PPAC trade-off in neuromorphic circuits with different current comparators (one of the critical analog components) by performing circuit simulation using our neuromorphic circuit optimization framework. Another objective is to study the performance of the circuit at cryogenic temperature (77K) for harsh environment applications. Participants in this project will: (1) learn how to use our optimization framework, (2) learn how to use a circuit simulator, (3) perform transistor measurements at 77K and calibrate models for cryogenic simulations, and (4) optimize the circuits.
Amplifying effects of electrical signals in circuit elements:
Stochastic resonance (SR) is a phenomenon where a signal that is normally too weak to be detected by a sensor, can be boosted by adding white noise to the signal, which contains a wide spectrum of frequencies. SR has been used to detect faint signals in bridge and mechanical failure analysis. Also in robotics, biology, neural modeling, and analog to digital conversion. The simplest system that demonstrates SR (and thus can be used as a detector) is a Schmitt trigger. A Schmitt trigger is an active circuit which converts an analog input signal to a digital output signal. The circuit is named a ”trigger” because the output retains its value until the input changes sufficiently to trigger a change. Schmitt trigger devices are typically used in signal conditioning applications to remove noise from signals used in digital circuits, particularly mechanical contact bounce in switches. They are also used in closed loop negative feedback configurations to implement relaxation oscillators, used in function generators and switching power supplies.
Objectives:
The first research objective of this project is to determine how the width of the hysteresis (the value of a physical property lags behind changes in the effect causing it) window of the Schmitt trigger affects the SR frequency range. The second objective is to determine how other operational amplifier parameters such as input offset voltage, slew rate limitations, gain bandwidth, and DC open-loop gain affects the SR parameters. The third objective will be to evaluate the ability of operational amplifiers designed by the researcher for silicon neurons to detect SR.