Testing predictive coding models
Our Centre research program on Prediction is based on recent demonstrations that the brain does not simply respond to external events, but rather compares sensory information against predictions based on internal representations (memories). The difference between predictions and external inputs (‘prediction errors’) are used to initiate adaptive behaviours. For example, when you are crossing the street, the observed trajectory of an oncoming car (sensory input) allows your brain to predict your movement relative to that of the car, based on past experience (memory) of trajectories of moving vehicles. Appropriate movements are then initiated to avoid collision. The computational load on the brain is thus reduced, from all-encompassing sensory perception to the more tractable problem of comparing sensory inputs to internally stored predictions. This ‘predictive error’ framework can be used to unify apparently diverse behavioural data, from low-level functions such as control of eye movements, through to attention and high-level functions such as decision.
Most of our physiologically based studies of Prediction employ the well-established fear conditioning paradigm in rodents. This paradigm is adapted to test predictive coding models by manipulating statistical regularities in the properties or timing of conditioned and unconditioned stimuli. For example, the probabilities have been changed so that only 50 per cent of the tones (conditioned stimuli) are paired with the shocks (unconditioned stimuli). Thus animals can modify their expectation of a shock and respond adaptively.
In other studies, we are using eye movement-based prediction tasks in non-human primates trained to ‘intercept’ visual targets based on their prior history of movement, with or without concurrent multisensory cues (e.g. an auditory stimulus that predicts stimulus acceleration or deceleration with different probabilities). This same paradigm can be adapted to human studies once we have learned more about the physiological signatures of prediction errors. The new approach that the Centre research brings is to undertake multi-scale experiments based on these paradigms, and analyse and interpret the data acquired from the scale of the single neuron, through neural circuits, up to the whole brain.