Career:The Adaptive Silicon Cochlea: Biology, VLSI, and Applications

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

We described a 117-stage 100Hz-to-10kHz silicon cochlea that attained a dynamic range of 61dB while dissipating 0.5mW of power. This cochlea has the widest dynamic range of any artificial cochlea built to date. The wide dynamic range was attained in a low-noise traveling-wave architecture with distributed gain control. The analysis of our electronic cochlea suggested why nature preferred a traveling-wave mechanism over a bank-of bandpass filters to decompose sounds. We propose to use the silicon cochlea in 3 projects that will range over a span of 4-5 years. The first project, Fundamental Issues in Cochlear Gain Control will extend and expand our previous work on gain control in the cochlea to improve its performance further, will help understand the functional role of gain control in the biological cochlea better, and will yield computational primitives and circuits that are useful for cochlear-implant speech processors, and for robust-and-adaptive speech and auditory-pattern recognition. We expect this project to be funded soley through the CAREER proposal. The second project, Cochlear-Implant Speech Processors for the Deaf will focuss on using the silicon cochlea as a low-power front end in cochlear-implant speech processors. We expect the performance of these processors to significantly improve because the silicon cochlea models several important effects that are exhibited by the biological cochlea that current implants do not, and also because it can implement these computationally-intensive biological algorithms in real time and with low power. The cochlear-implant project will be done in collaboration with Dr. Donald Eddington at the Massachusetts Eye and Ear Infirmary, who is a pioneer in the field of cochlear implants and who is very happy to collaborate with us. We expect to use CAREER funding to help us get started on this project, but fund it primarily via other grants from the NIH or NSF. The third project, Adaptive Front Ends for Speech and Pattern Recognition will focuss on using spike-based hybrid (analog-digital) computation techniques that were invented by the author at Bell Labs to construct an energy-efficient and adaptive trellis vector quantizer, an important component of speech preprocessors. The trellis vector quantizer is useful in quantizing sequences of analog vectors with certain temporal dependencies between them, such as those that emerge from a cochlea; we plan to implement onchip learning to adapt the parameters of this quantizer; the architecture of such a quantizer maps naturally to the architecture of a hybrid state machine (HSM), a machine that extends and expands the concept of a finite state machine to the hybrid domain. We expect to supplement the beginning funding for this project from the CAREER proposal with funding from Lucent Technologies or other funding agencies. We will collaborate with speech researchers at Bell Labs and at MIT on this project. We propose to develop two courses at MIT, entitled Feedback in Electronics and Biology, and Hybrid Computation over a span of 3-4 years that will introduce into electronics, neural examples and concepts in adaptive-and-hybrid computing.