spatial computation in the entorhinal cortex

In my PhD project I am trying to get a better idea of how our brains perform spatial computations. Our lab looks at the medial entorhinal cortex (MEC) in particular, a brain region known for neurons whose activity correlates with spatial variables. One striking example is the grid cell, which fires in multiple locations that form a hexagonal grid (as you might have noticed in the background).

Most models of spatial computation in the MEC are built by wiring up these neuron types in various ways. In my first year, I hypothesised that if these functionally defined neurons are indeed the building blocks of spatial computation, then their functional identity should be stable across spatial tasks.

I quantified this stability across an open field foraging and a virtual reality navigation task, and found that there exist neurons with stable identities, but there are also some that change their identity depending on the task. Particularly neurons whose firing is correlated with the animal's speed seem unstable:

foundation models of primary visual cortex

A project led by Bryan Li, who I am very grateful to be able to work with.

The goal here is to leverage machine learning models to extract patterns and learn from the increasingly large neural recordings that are being collected in the primary visual cortex (V1) of mice. We designed a spatiotemporal transformer model and trained it to predict the activity (2-photon calcium imaging) of over 8000 V1 neurons from 10 mice. In terms of pure predictive power we managed to get third place in the Sensorium 2023 competition, earning us some real estate in the retrospective NeurIPS paper (accepted at NeurIPS2024).

However, aside from pure predictive power, we are working on showing that these types of models learn computational principles in mouse V1, allowing them to extrapolate and thus making them great hypothesis generators:

As outlined in the paper from the Andreas Tolias lab on "Inception Loops", we are also looking at how these models can be used to generate stimuli that are particularly interesting to test on real animals, thereby reducing in vivo experiments.

models of sniff- and odor-driven responses in the olfactory cortex

I was very fortunate to be a NeuroAI intern in the Albeanu lab at Cold Spring Harbor Laboratory, New York. During this project I designed Poisson Generalised Linear Models (GLMs) to predict the spiking of olfactory cortex neurons from sniff and odor features.

Neurons in the olfactory cortex are strongly modulated by sniff cycles, making it very hard to disentagle the contributions of any other stimulus with the sniff-driven responses. With these simple and interpretable single neuron Poisson GLMs, we hope to be able to get a better understanding of what the olfactory cortex is actually doing.

Figure coming soon!