Research

High Throughput Connectomics

Connectomics is a burgeoning field within neuroscience that attempts to better understand mechanisms of the brain through tracing the exact connections within it. While there are many scales in which connectomics can be employed, we specifically are interested in analyzing the exact synaptic connections between individual neurons. Doing so will allow for a better understanding of the mechanisms of medium scale neuron circuits. At our lab, we are collaborating with the Lichtman Lab at Harvard University to speed up the collection of connectomics datasets massively as part of the Smart-EM project. We are also analyzing the MICrONS connectomics dataset to understand how the mouse visual system is organized.

Efficient Machine Learning

As large language models continue to grow in capability, so too do they grow in size. Such a blowup has significant environmental and monetary cost, both for training and inference. We have discovered the existence of interesting neurons within LLMs that are multimodal and have an outsized sensitivity to pruning. Additionally, our lab closely collaborates with other researchers in the space, such as the Dan Alistarh Group at IST Austria, to develop new ways to compress LLMs post-training for less computationally costly inference. We are also improving training performance through efficient neural network growth.