Skip to main content

Bayesian inference in neural dynamics: linking biophysical and computational approaches to neuroscience

Final Activity Report Summary - BIND (Bayesian Inference in neural dynamics: Linking biophysical and computational approaches to neuroscience.)

Recent years have seen the growing use of normative Bayesian models to describe behaviour, perception and reasoning. These models formalise sensory perception, motor control or behavioural strategies as probabilistic inference and learning in elementary causal models. The goal of the 'Bayesian inference in neural dynamics' (BIND) project was to adapt this powerful set of tools to further our understanding of the function and dynamics of biological neurons and neural networks, and how these underlie cognition.

Using this approach, we were able to assign precise functional roles to basic biophysical entities such as single dendrites, single neurons, synapses and small circuits. For example, the time constant of single neurons and synaptic dynamics can implement prior knowledge about stimulus probability and duration, and be adjusted by local spike-based learning rules. Similarly, neural networks can implement models of the outside world and be tuned to detect important sensory events optimally.

"Normative models" suppose that neurons are primarily predictors of their inputs. Neural networks learn and infer which sets of events in the outside world most probably caused the sensory inputs. We showed that this explains why the responses or sensory neuron are so highly adaptive and context dependent, and so poorly described by their receptive fields. On the other hand, we showed that these adaptations and contextual interactions are captured by a very simple normative model. This approach could radically change the way we think of sensory neural representations by replacing the concept of "receptive fields" by the symmetrical concept of a "predictive field".

Similarly, we showed that the activities of population of neurons can be understood as representing, computing and memorising the probability distribution of dynamic perceptual and behavioural variables, such as direction of motion, behavioural choices or eye movements. The integrate and fire dynamics of cortical neurons imply that each spike represents a sudden increase in probability, or, equivalently, a prediction error between what the "brain already knows" and the new sensory input. If we were able to perfectly predict the world, there would be no spikes at all. As a consequence, the apparent variability of neural responses (for example the fact that spikes never occur at the same time on repeated trials) masks extremely precise computations. The brain is not as noisy as it seems, and response variability reflects the uncertainty associated with each perceptual interpretation of the sensory input.

We coupled this normative "top down approach" with a "bottom up approach" that starts from what we know of the dynamics and biophysics of single neurons and networks and infer which tasks they might be implementing. We identified currents recruited in single neurons that can make them function as Bayesian integrator. We studied which time based computations can and cannot be implemented by active dendrites. We identified neural correlates of decision variables in a tactile discrimination task by analysing the responses of hundreds of neurons recorded previously in primary tactile areas and premotor areas of primates. We modelled the impact of nicotine on different receptor subtypes in the ventral tegmental areas and showed how its affect on network dynamics underlie drug addiction. This led us to propose a new role of acetylcholine in coding prediction errors in reinforcement learning. The picture drawn from this bottom-up approach is multifaceted and illustrates the richness and complexity of the neural dynamics involved in even basic computations. At the same time, the biophysics of the nervous system strongly constrains normal and pathological behavioural functions, bringing us closer to understanding how belief systems can be embodied in a neural substrate.