by Federico D'Agostino - Submitted to University College London on 12/09/2022
Download the full PDF with Appendix material here.
<aside> <img src="https://upload.wikimedia.org/wikipedia/commons/c/ca/1x1.png" alt="https://upload.wikimedia.org/wikipedia/commons/c/ca/1x1.png" width="40px" /> The advent of revolutionary new multisite silicon probes equips researchers with tools to record simultaneously from an unprecedented number of neurons, opening new avenues for the description of neural circuits. However, numerous molecular, morphological, functional and connectional characteristics distinguish neurons in the brain into different cell types, and reliable identification of such cell types in extracellular recordings is essential to determine their contributions to neural circuit computations. This is difficult in any brain area, but particularly challenging in the cerebellar cortex due to the high density of neurons, their high firing rates, and the elaborately folded cytoarchitecture. Previous studies tried to solve this problem using supervised learning, but lacked rigorousness in their machine learning pipelines and used datasets coming from anaesthetised animals. Here we tackle the problem using a novel dataset coming from high-density Neuropixels recordings of the cerebellar cortex in awake, freely moving mice. Due to the complexity of the experimental protocol for data acquisition, only a small amount of the data is labelled. As a consequence, for the first time in this domain, we adopt a range of modern deep semi-supervised methods to approach the task, making the most efficient possible use of ground-truth information. Results show how our models are able to surpass in accuracy both human experts and a baseline constructed with engineered features from the electrophysiology literature, in some cases using just only a fraction of the total labels available. We further propose concrete steps to bring our model architectures into deployment, to yield a tool that can be reliably incorporated into the analysis pipelines of electrophysiology laboratories across the world. Our broader hope is to inspire researchers in biology to make a more resource-aware use of data, especially when coming from costly and time-consuming experiments.
</aside>
