Banner image representing models
Prev Next

Generative models, inference and predictions

Sensory and abstract representations are brought together to form generative models of our environment, which are updated with external and internal information to allow us to predict the world around us. At SWC, researchers focus on understanding how these models are represented in neural activity and how the brain learns and stores prior knowledge (‘priors’).

A key assumption of the generative model is that it can be used to influence sensory processing to predict, disambiguate or prioritise sensory input. Previous research at SWC revealed that sensory representations not only reflect the identity of external stimuli but also contain information about stimulus context, behavioural relevance as well as current goals, expectations, actions or past experience. 

Our latest research is revealing how diverse contextual signals are integrated with sensory information, and how such signals emerge through experience and learning.

Current research

The Akrami lab is using behavioural and physiological approaches, along with computational modelling, to study the neural underpinnings of the formation, storage and utilisation of experience-dependent priors. The team focuses on understanding how probabilities are represented in the brain and is using this knowledge to develop new circuit-level models of the computations that underlie formation of priors. They are also testing how statistical learning of regularities in sensory inputs or abstract representations can be stored in neural circuits.

The Saxe lab is exploring how prior knowledge impacts learning. The team are investigating the dynamics of learning new tasks with a neural network that already has pre-existing knowledge.

The Hofer, Margrie and Mrsic-Flogel labs are working to understand the underlying circuit mechanisms in visually-guided behaviour. They use anatomical tracing to identify candidate regions and then measure and disrupt feedback from identified regions during visually-guided behaviours. Furthermore they are testing how these inputs influence the local network activity in relation to different excitatory and inhibitory cell classes in visual cortex, and are working with GCNU to build a model of how long-range predictive signals influence local computations during sensory processing.

The Murray and Mrsic-Flogel labs are investigating how animals adapt to deviations from expected sensory feedback. They work on the cerebellum, which is known to store the internal model of sensorimotor transformations required for movement and also learns to control the timing of actions in response to predictable changes in sensory input. They have shown that cerebellar output is crucial for generating activity dynamics in premotor cortex during planning and they are now investigating the contributions of the cerebellum to sensorimotor representations in the neocortex as animals adapt to deviations from expected sensory feedback.

The Erlich lab is working closely with the Duan lab to understand social cognition and how the brain predicts actions of other agents in the world. While technical limitations have meant that systems neuroscience has traditionally focused on studying individual animals, the Erlich and Duan labs are making use of recent advances in computer vision to track multiple animals in real-time and extract detailed information about their behaviour. 

Additionally, the Erlich lab is working to understand how information that comes into our nervous system in a sensory reference frame is used to update a stable world-model that enables goal-direction behaviour.