Keywords: Ca2+ imaging, mouse, behavior, neural, manifold, learning
Project coordinator: Prof. Flavio Donato
CeDA collaborator: Rodrigo C. G. Pena

Context

The brain's ability to support sophisticated cognitive function like learning, memory, navigation, and decision making relies on the interaction of thousands of neurons that change their activity over time according to dynamics that are ultimately related to what an animal is experiencing on a moment-to-moment basis. At the level of individual neurons, activity dynamics integrate signals coming from the sensory world with self-generated activity patterns that largely result from the topology of the network in which these neurons are embedded. At the neuronal network level, the temporal progression of such population dynamics keeps track of:

  • the temporal unfolding of sensory stimuli impinging on the animal,
  • the progressive changes in behaviorally-relevant variables like, for example, the animal's position in space or distance to goals,
  • dynamic changes in the animal's internal states, and
  • an internal model that predicts the outcome of the animal's actions.

In the past years it has become evident therefore, that to understand how populations of neurons support cognitive functions it is increasingly important to describe emergent phenomena that modulate the propagation of activity in biological neuronal networks, and how they relate to an animal's behavior. Lately, a successful framework for the study of such emergent properties has been the use of manifold-learning algorithms to define a complex, yet low dimensional, network space where neuronal activity resides, discover the topology of such space to infer connectivity rules between neurons, and reveal the dynamics that characterize the unfolding of neuronal activity on the manifold over time.

Project objectives

The aim of this collaboration is to develop novel strategy to reveal and align the intrinsic manifold of neural activity to the low-dimensional space describing task variables, in order to understand how activity dynamics generated in a classical autoassociative network like the one instantiated in the rodent hippocampus is able to represent information about an animal's experience, to extract general rules about complex tasks and learn from past episodes, and to develop the ability to produce inferences about future states of the network in order to take decisions and plan actions.

Approaches

  • Learn lower-dimensional representations of the neural activity of a mouse performing a sequence of actions.
  • Train auto-encoder-like neural networks to reconstruct calcium (Ca2+) imaging of neural activity.
  • Study the shape of the trained auto-encoder's latent representations and check how they relate to behavior.