In this era of AI, often deep neural networks (DNNs) are viewed as an “all-cure” solution; their success in a wide range of problems needs no mention. Unfortunately, such systems are generally black boxes and somewhat related to how biological neural networks (brains) which are claimed to develop “intelligence”. Although it is hard to define “intelligence”, achieving human-level performance (or even better) in different decision-making problems using systems with billions/trillions of free parameters trained on huge data sets may not necessarily make a system intelligent! In this context, we shall discuss how the design of some of the AI systems used inspiration from neuroscience knowingly or unknowingly. We argue that to realize human-like intelligence a closer interaction between discoveries in neuroscience and the designing of AI systems is needed. However, we need to acknowledge that the brain is probably the most complex object in the known world with more unknowns than knowns, although we know a lot about it. Truly brain-inspired models may be able to put a brake on the apparently unsustainable philosophy of “bigger the better” – bigger architecture and bigger datasets. We shall then discuss some of our attempts to exploit neuroscience models/discoveries to develop pattern-recognition systems (intentionally we are avoiding the term AI). Particularly, we shall demonstrate that exploiting at a high level some findings from a cat’s visual cortex can make a multilayer perceptron a bit more comprehensible. We shall also discuss how computational models of cells like Lateral Geniculate Nucleus (LGN) cells and Retinal Ganglion Cells can be used to extract features from images to use as inputs to CNNs to investigate if it can improve the complexity and performance of the system. The answers turn out to be affirmative.

Speaker: Professor Nikhil R. PAL
Date: 9 May 2024 (Thursday)
Time: 10:00am – 11:00am
Poster: Click here