The Bayesian Brain
The Bayesian Brain
The Bayesian brain considers the brain as a statistical organ of hierarchical inference that predicts current and future events on the basis of past experience. According to this theory, the mind makes sense of the world by assigning probabilities to hypotheses that best explain (usually sparse and ambiguous) sensory data – and continually updating these hypotheses according to standard probabilistic rules of inference.
This fine-tuning (optimization) of perception and action operates under the single imperative of minimizing surprise (free energy) and uncertainty; thereby maximizing statistical and thermodynamic efficiency.
Learning in the Bayesian brain differs from reinforcement (and machine) learning because it occurs with understanding. Mental models of past experience use these experiences to anticipate new experiences, as opposed to being shaped by them. Continual optimisation of the models also enables efficient exchange with the environment in a self-organised, self-evidencing and unsupervised fashion.
We use neuroimaging to investigate the central tenets of hierarchical inference in the brain:
- Brain regions within a neural system are hierarchically organised.
- Each level of the hierarchy sends top-down signals to predict inputs at lower levels.
- Prediction error = information at a lower level that is not predicted by a higher level.
- Higher levels update their predictions to minimise subsequent prediction error.
- The degree to which higher levels update depends on the precision of prediction error
- Precision (reliability) in the prediction error is estimated from prior experience.
- The percept is dictated by the synthesis of new sensory information with prior beliefs.
- Attention increases the precision of prediction errors and their influence on high-levels.
- Predicted actions are realised when the precision of prediction errors is attenuated.
- Bottom-up prediction errors are thought to arise at 30-70Hz from superficial cortical layers.
- Top-down predictions are thought to arise at 20-30 Hz from deeper cortical layers.
- Top-down predictions of precision are mediated by neuromodulatory mechanisms.
- Behaviour (e.g. actions) minimises uncertainty (expected surprise). This is known as ’active inference’.
For example, Dynamic Causal Modelling (DCM) of fMRI, MEG and EEG data can be used to:
- distinguish forward and backward neuronal signalling in different layers and levels of the cortex,
- specify the mechanisms of neuronal communication at the synaptic and neurotransmitter level and
- elucidate the principles of inference in the brain and its role in psychopathology.
Our goal is to provide:
Mathematical models of computations, microcircuits and neuronal processing that can be used to develop and validate a unifying framework for understanding:
- normal and abnormal perception
- decision making
- many other functions.
Neurological and psychiatric disorders may be caused by problems updating hierarchical predictions in the brain – at the synaptic and neurotransmitter levels; particularly in the neuromodulatory control of precision.
This could result in seizure activity in epilepsy, difficulty interpreting new information on the basis of past experience (false inference) and difficulty updating expectations about the body or environment.
Examples of false inference include
- hallucinations and delusions in schizophrenia
- failure to infer motor states in Parkinson’s disease.