Began our work to decode invasive neuron spiking data for image category classification

Keywords: Neuroscience + AI, Invasive Neuron Decoding, Data Constraints

We had a groundbreaking dataset of signals recorded straight from the hippocampus and amygdala using invasive electrodes. The challenge? These signals were recorded with the electrodes inside the patients’ heads, in real-time, viewing memory-related tasks. Such a risk was only dared to be taken on by 42 patients who have had significant trouble locating their epilepsy before. This experiment was their last resort. With such a small dataset, overfitting posed as a significant possibility and challenge. We hacked away a this task under the guidance of our graduate student mentor Geeling Chau (advised by Prof. Yisong Yue) and Prof. Katie Bouman, working to parse through the sparse set of neuron spike data for signals.

The goal was to classify these neuron spikes to recover which category of images the patient was viewing at the time. For half the term, we were stuck at a near-chance level of accuracy in the classification (out of 5 categories). Only after many, many attempts and combining everything we’ve tried (binning the spikes in various ways, grouping various neurons as a category rather than all at once, running different recurrent models, etc), were we able to nearly triple the accuracy while driving down the overfitting. Still, the results varied heavily between patients. This was the project that began my research journey, and definitely one of the projects that taught me perseverance and patience the most 😂