Oct 26, 2016
Jolley Hall, Room 309
"Analog Inside: Nonlinear System Accelerators & Ultra-low-power Sensors"
Advisor: Shantanu Chakrabartty
In this proposal, I explore two new frontiers in analog computing that may yield highly optimized and energy-efficient designs for emulating nonlinear systems and for ultra-low-power sensing.
In the first case study, I consider the "jump resonance" phenomena, which is observed in nonlinear circuits when the amplitude of the output signal exhibits an abrupt change for a well-behaved periodic input signal. For analog filters used in auditory front-ends, this nonlinearity is treated as an artifact that needs to be alleviated; however, surveying naturally occurring auditory systems reveals that evolution has yielded a biological complex containing a myriad of nonlinearities. I posit that using this phenomena will enable the extraction of more discriminative features to improve speaker recognition performance. Preliminary results validating this hypothesis will be presented.
In the second case study, I explore the physics of hot-electron injection to design an ultra-low-power microsystem to: sense, compute, and store statistics of infrasonic signals. Infrasonic signals are those with frequencies below 20 Hz, with physical signal sources like ambient temperature swings, pressure variations, or structural accelerations. I propose to investigate a novel paradigm for designing an ultra-low-power sensor with filtering and data-logging that leverages the physics of hot-electron injection to implement large system time-constants, signal rectification, and non-volatile storage. Doing so will obviate the need for large signal transducers whilst enabling the monitoring of spectral features well below 20 Hz.