Johns Hopkins School of Medicine

As we explore and navigate the world, we are constantly learning about our environment -- for example we implicitly learn the visual features on a road that we traverse each day to work. These visual features are high-dimensional, yet somehow they are compressed into lower dimensional abstract features that we can recall. Our lab is interested in understanding how the brain creates visual memories, interrogating the plasticity rules of such learning, and determining how animals generalize to new environments using these learned features. We also study the representations of motor actions in cortical circuits and are interested in how these representations may play a role in sensorimotor learning.

 

To answer these questions, we quantify neural activity in visual areas before, during and after learning. For this we use two-photon calcium imaging, and record over 50,000 neurons simultaneously across several cortical areas across weeks in mice traversing complex closed-loop virtual reality corridors. We observe considerable plasticity in visual circuits, even when the mouse is not receiving rewards, i.e. unsupervised learning. There are several different algorithms for unsupervised learning proposed in theoretical neuroscience and deep learning. Our lab takes inspiration from these theories and tests them via careful experimental design and computational analysis of the neural activity. This computational analysis often requires the design of new modeling techniques due to the high-dimensional nature of the neural activity that we record.