Neural Nets as Scientific Instruments
Predicting neural activity or behaviour is not enough. We seek mechanistic, process level explanations of cognition.
Sushrut will argue that modern neural networks can be viewed as scientific instruments for theory-building: by intervening on architecture, training diet, and objective, we can test which changes reproduce or break cognitive and neural phenomena.
He'll illustrate this through two case studies - how a human inspired developmental visual diet, capturing changes in acuity, contrast sensitivity, and colour vision over time, leads to shape-biased vision; and how representational drift connects to continual learning, where systems maintain performance on older tasks as they acquire new ones.
The core takeaway: Performant Neural Networks don't automatically explain cognition, but used as controlled experimental systems, they can help us construct and validate theories of brain and behaviour.
Speaker Bio
Dr. Sushrut Thorat is a cognitive neuroscientist / ML researcher who uses neural networks as controlled model systems to study perception and learning.
His work spans human-inspired developmental vision curricula, scene perception, cognitive behaviours, and continual learning theory, to link model mechanisms to cognitive and neural phenomena.
Pre Read
Optional skim (abstract-level is enough):