UCL Queen Square Institute of Neurology


Uncertainty and inference

30 November 2017

New research by Thomas Parr and Karl Friston, UCL Institute of Neurology, draws upon recent advances in theoretical neurobiology to try to understand how the brain might solve the problem of dealing with environmental uncertainty.

They asked what optimal exploratory behaviour would look like when the world changes in an unpredictable (or ‘volatile’) way, or if it gives rise to unreliable (‘imprecise’) sensory information.

To answer these questions, they simulated the (exploratory) eye movements that a synthetic brain would make in a simple visual environment, equipping this brain with explicit beliefs about the uncertainty in different locations. They found that several well-described psychological phenomena emerged from the simulation. The first of these is ‘inhibition of return’. This is the tendency not to look back to a particular location when recently having looked at it. Most computational models of this effect write this in explicitly.

"We found that it is an emergent property of a system that holds beliefs about the volatility of its world – intuitively, if things in the world are changing in an unpredictable way, the longer it has been since looking at something, the more information there is to be gained by looking at it. Conversely, if it has been seen recently, it is more useful to look at other things that have not been seen for a while." Thomas Parr

The second phenomenon exhibited by their in silico brain is the ‘streetlight’ effect. This is often given as an example of an (unhelpful) cognitive bias that involves searching in places that we can already see. The name comes from the example of a drunkard searching for his keys under the streetlight, as it is the only place he can see. Counterintuitively, we found that this kind of behaviour is actually very sensible. In order to effectively gather information, it is best to look more frequently towards locations that yield high quality visual data. Our synthetic agent dutifully ignored locations that provided little useful information (like those outside the range of the streetlight).

In addition to demonstrating the emergence of these interesting behaviours, they asked how the brain might work out the levels of uncertainty in the environment. In answering this question, they derived the form of the neural networks that could perform these computations. Interestingly, these mapped closely to the anatomical connections of the brain systems thought to be involved in computing uncertainty.

"This research complements and builds upon previous theories that noradrenaline, acetylcholine, and dopamine are involved in signalling uncertainty. It additionally allows us to think of disorders of these systems as failures to correctly estimate the volatility of the world, and the reliability of the sensations it generates. Similarly, we can start to think about drugs that act upon these systems (e.g. donepezil for Alzheimer’s disease, and the noradrenaline uptake inhibitors used to treat depression) in terms of the aberrant neural computations they try to correct." Thomas Parr

Further information

Image: The neural networks required to compute uncertainty, and their correspondence to the cholinergic (upper image) and noradrenergic (lower image) systems.