Mock Bredenberg Lab Research

While many models exist that seek to describe how the brain could learn to build sensory representations or learn from reinforcement using synaptic plasticity, most of these models do not perform nearly as well as state-of-the-art machine learning methods.

We are working improve the performance and realism of biological learning models by to extracting computational motifs from our knowledge of low-level cortical information processing, to incorporate into neural networks whose architecture will synergize effectively with models of biological learning. 

In recent years, numerous statistical tools have been developed to jointly characterize neural representations and behavior in terms of low-dimensional manifolds and dynamics; however, there is a comparative absence of similar techniques that describe learning-induced changes in the brain.

Our lab is working to create approaches to statistically quantify learning in neural representations and behavior. 

To do this, we collaborate with experimental neuroscientists to develop novel statistical methods to characterize the neural basis of learning in naturalistic environments, using a spectrum of approaches, from simple, interpretable methods to powerful deep learning-driven decoding approaches, ranging from linear tensor decompositions, to dynamical similarity metrics, to hierarchical, nonlinear state-space methods.

We aim to generate testable predictions that can distinguish candidate models from their alternatives. To do this, we have collaboration with Allen Institute for Brain Science to test candidate models for how mice learn to control brain-computer interfaces (BCIs). We are also collaborating with experimental neuroscientists to explore whether underlying classical psychedelic-induced hallucinations and dream generation are one and the same.