Abstract:

Developing abstract models to emulate the functionality of low-level
complex mechanisms has contributed significantly towards several advancements in computer architecture. These abstract models are relatively easy to implement, they are computationally efficient, and they allow the architects to study interactions among high-level modules while avoiding the intricacies and implementation details of low-level constructs. A similar approach has been adopted in neural network design; however, within this paradigm the high-level abstract
models do not fully capture the key aspects of their biological inspirations. As a result, these high-level neural abstractions significantly compromise the power of the underlying biological architecture.

This talk describes cortical columns as one of the high-level
biologically inspired computational abstractions. It provides insights into the biological basis of the column model and demonstrates the advantages of such a neural abstraction. Afterwards, it introduces spiking neural networks as an interesting low-level neural abstraction and describes various powerful mechanisms provided by spiking neurons that are not captured by the aforementioned column model. This
disconnection between the high and low level neural models severely limits the applicability of high-level neural models towards performing complex tasks like perception, attention, and decision-making. High-level neural models have their advantages; however, to properly develop such models, their low-level biological inspirations must be studied and modeled in details. Only then can truly intelligent autonomous systems be developed. Computer architects are uniquely
positioned for this task because, unlike biologists, their goal is to steer this research towards useful computing systems and applications. Although both the elementary components and the resulting biologically-inspired systems are quite different from existing computational hardware, similar architectural approaches can and should be used to model these biological abstractions.

Biography:

Dr. Hashmi obtained his M.S. and Ph.D. in Electrical
Engineering from the University of Wisconsin (UW) – Madison. His research focuses on understanding structural and functional properties of the mammalian brain and developing computational models inspired by these properties. Presently, Dr. Hashmi is a post-doctoral research associate in the Sleep and Consciousness labs at UW – Madison working with Professor Giulio Tononi and is studying complex neuronal dynamics and their role in perception, attention, and decision-making. He is also part of the UW – Madison research team collaborating with IBM on the SyNAPSE project.