Artificial Intelligence Colloquium - Celestine Lawrence PhD University of Groningen

Title: An equivalence relation for state-space models : from neurophysiology to nanoelectronic networks for realizing cognitive systems

Abstract:

This talk will begin with a historical overview of developments in neurophysiology since the 17th century and reflect upon the concepts that have been adopted (or abandoned) by the machine learning community. Based on these neurophysiological insights and the constraints of real-world signals, we will argue that a resource-efficient realization of an octet of signal processing primitives is necessary for achieving cognition, namely : concentration invariance, uniform time-warp invariance, sequence generation, pattern recognition, signal transformation, self-organized feature maps, independent component analysis, and nonlinear blind source separation. We will make it mathematically evident that there exists an equivalence class of state-space models to realize these signal processing primitives. In other words, a systems biologist would say that there exists organisms (natural or artificial) of identical cognitive capacity but with different dynomes (portmanteau from dynamics and connectome). A major motivation for the above study is to eventually engineer unconventional cognitive systems based on 21st century developments in nanoelectronics. It is hoped that such nanoelectronic systems can achieve cognition with dynomes of higher dynamics to compensate for their lesser connectivity in comparison to neurophysiological dynomes. A demonstration is yet to be made, but it is postulated that spatially delineated modules of working memory, procedural memory, declarative memory, sensory input and motor output as in “computational” cognitive models would be realized as spatiotemporal modules of signal processing.