Towards Low-Power On-chip Auditory Processing

Abstract

Machine perception is a difficult problem both from a practical or implementation point of view as well as from a theoretical or algorithm point of view. Machine perception systems based on biological perception systems show great promise in many areas but they often have processing requirements and/or data flow requirements that are difficult to implement, especially in small or low-power systems. We propose a system design approach that makes it possible to implement complex functionality using cooperative analog-digital signal processing to lower-power requirements dramatically over digital-only systems, as well as provide an architecture facilitating the development of biologically motivated perception systems. We show the architecture and application development approach. We also present several reference systems for speech recognition, noise suppression, and audio classification

Finding a roadmap to achieve large neuromorphic hardware systems

“…A lot of previous work has focused on front-end sensory and motor systems, including retina models (e.g., Mead, 1989; Boahen and Andreou, 1992; Delbruck and Mead, 1994; Delbruck, 1994; Marwick and Andreou, 2006; Lichtsteiner et al, 2008) cochlea models (e.g., Mead, 1989; van Schaik et al, 1996; Sarpeshkar et al, 1998, 2005a,b; Ravindran et al, 2005; Hamilton et al, 2008; Odame and Hasler, 2008; Rumberg et al, 2008; van Schaik et al, 2010; Rumberg and Graham, 2012) as well as others (e.g., LeMoncheck, 1992). Although these input representations are important for neural computation, and some have done some interesting engineering work based on these front-end systems (Riesenhuber and Poggio, 2000; Fu et al, 2008; Schaik et al, 2009; Chakrabartty and Liu, 2010; Liu and Delbruck, 2010; Farabet et al, 2011; Sejnowski and Delbruck, 2012), our focus will be on the computation using these front-end structures in the highly modular cortical structure (Eliasmith and Anderson, 2003).…”

Reconfigurable analog signal processing for wireless sensor networks

“…It is a compelling choice to use analog processing in sensor networks due to the analog nature of real-world signals and due to the 20 year leap that analog electronics have been shown to have over digital systems in terms of performance-per-power consumed [4].…”

Data Conversion Within Energy Constrained Environments

“…Since the signal is analog to begin with, analog electronics offer a natural and power efficient means of doing so [22]. In [23], for example, an analog auditory sensory system was presented that provided a power savings of 3-4 orders of magnitude over a comparable digital system. This power savings was shown to be the equivalent of a 20 year leap in digital fabrication process technology.…”

“…Analog computation and pre-processing has been used in a wide variety of systems to improve energy savings, showing in some cases the equivalent of a 20-year leap in digital scaling [77]. Traditional analog pre-processing stages tend to be highly specialized application-specific systems, but developments in reconfigurable field-programmable analog arrays (FPAAs) [26,78] have allowed these analog techniques to be applied to systems without a priori knowledge of the application space.…”