Site logo

Brain & Behaviour lab

Dept. of Bioengineering & Dept. of Computing - Imperial College London Contact Me
Stacks Image 75
Stacks Image 77
At the computational level we study  the problem of sensorimotor learning and control: The only way the brain can interact with the world is through movement (think about it for a moment). Sensorimotor control is a non-trivial problem, given that we have to orchestrate in real-time hundreds of muscles acting over many redundant degrees of freedom, sensory & perceptual uncertainty, motor variability and external constraints set by having to act in a dynamic world.

What computational principles should (and does) the brain use to control movements? To investigate this we carry out psychophysics experiments using virtual reality interfaces and haptic, robotic devices. These experiments are driven by computational theories that combine ideas from Machine Learning (Bayesian Decision Theory) and Engineering (Stochastic/Optimal Control Theory). We translate these  normative theories into directly testable experimental tasks, to the point that we can predict each individual’s subject performance (
Faisal & Wolpert, 2009, J.Neurophys).
Ultimately, we want to know how any computational principles that are effective in sensorimotor control (i.e. they are proven to predict behaviour) have to be implemented in neural circuits. We take the view that constraints on the neural circuit level (
How fast, costly and reliable are my neurons?) and the behavioral level (How precise need our movements be?) are going to determine how neural circuits and behaviour interplay.
Stacks Image 79
At the Implementation level we study a fundamental aspect of the nervous system: Signaling and noise in nerve fibers. Translated into computer science: how reliable and how miniaturized is the brain  given that it has wires made of salty water insulated by fat and transistors made of single proteins? We showed that noise in these protein transistors (so called voltage-gated ion channels) sets a universal lower limit to the diameter of axons (size of cells) across species at 100 nm diameter and that it will affect signalling reliability in the many thin fibers of our brains (e.g. see Faisal et al, 2005, Curr Biol; Faisal & Laughlin, 2007, PLOS Comp Biol). Our brain cells respond a million times slower than current CPUs, it consumes only 17 Watts and yet we can cycle or read this text in “real-time” - we are interested in uncovering the principles that enable neural circuits to achieve this.
To support our scientific investigation we developed stochastic integration algorithms and data structures for partial differential equations and implemented the Modigliani stochastic simulation framework (
www.modigliani.co.uk) for biophysically detailed simulations of neurons and neuronal circuits.

Stacks Image 83
Bioinformatics of Behavior - Linked to our neuroscience research program we are developing technology for neuroscience (e.g. Faisal & Matheson, 2001, J.Exp.Biol; Faisal &; Niven, 2008, LNCS; Faisal, Rogers & Niven, 2008). What can the study of behavior tell us about the neural circuits and genes that generated it? Ethomics - the systematic study of links between behavior and genes (which determine how neuronal networks are assembled together) is ramping up internationally. Traditionally, variability in behavior has complicated it’s study. While in the 20th century we already quantified the activity of cells and genes with great precision, at the beginning of the 21st century behavior is often still studied “by eye”.

This motivates the development of our framework, the Bioinformatics of Behaviour. Our analysis approach is to preserve and exploit the fact that behavior is variable, often containing meaningful information about the underlying mechanisms. We are currently applying Hidden Markov Models to detect meaningful structure in Human toolmaking and Insect behavior in complex tasks.
Our data gathering approach is to adapt hardware used in human movement studies and computer vision. Ultimatley we combine behavioural data gathering and analysis by using with machine learning methods to automatically annotate,  analyze and extract structure in complex behavioral data of humans and animals with defined genetic or electrophysiological perturbations.