Neuroscience research is a vital source of inspiration for the design of Artificial Intelligence (AI) systems. Whilst, ultimately there is no reason that biological networks represent the only successful blueprint of a cognitive system, at this time they are the best available model. Accordingly, detailed understanding of the primate visual system has framed design for artificial visual systems such as Convolutional Neural Networks (CNNs) and Recursive Cortical Networks (RCNs), and understanding of mammalian memory systems have inspired Long Short-Term Memory (LSTM) and Deep Reinforcement Learning (Deep RL) systems.
Nevertheless, despite significant advances in AI and cognitive neuroscience research, we are still some way from developing human level AI. Whilst partly due to engineering constraints – human brains are vastly complex systems capable of performing multiple parallel processes, interpreting, generalisation and making informed decisions based on inference from multiple concurrent sources of sensorial information; a significant constraint is our lack of complete understanding of how human brain’s work, particularly with respect to higher order cognitive systems.
In this recent talk, given for the Machine Learning Tutorial series, organised br Dr Marc Deisenroth, from the Department of Computing Imperial College, I argue that part of the reason for this is related to the challenges of localising these higher order systems, consistently, across broad populations of brains. I present the potential and challenges involved with learning models of the brain using Magnetic Resonance Imaging (MRI). I provide practical guidance to the techniques used to maximise the accuracy of these analysis through machine learning based data clean-up, discuss state of the art methods for modelling whole brain connectivity networks, and propose new techniques for improving the sensitivity of population-based analysis through the use of Deep Learning.
The slides for the talk can be found here