Sheaves are an incredible interesting concept in mathematics and there is a lot of potential applications for them in science. Unfortunately, the mathematical details are extremely dense and it is hard to gain an intuition for what's actually going on. So I wrote a paper doing my best to explain them in a very elementary way with lots of concrete examples and anaologies instead of the one-sentence definitions only intelligible by people who already know what a sheaf is. close

A long standing hypothesis is that premotor cortex plans movement and sends those signals to the primary motor cortex which executes those movements. This is simplifying things greatly, but that is what we set out to investigate using new the technology available in Neuropixel Probes, which can measure from hundreds of neurons simultaneously. By recording neural activity in primary motor cortex and optogenetically inactiviting neural activity in premotor cortex, as well as vice versa, in a mouse doing a skilled reaching task, we were able to see how this hypothesis held at a level of granularity never achieved before. Consistent with the idea of hierarchy, the premotor cortex had a much stronger influence on primary motor cortex than the reverse direction.

This was exciting and so the next step was to record from both of these regions simultaneously and use functional connectivity metrics to measure the flow of information between the two regions in as many ways as possible, to make sure that we weren't missing anything. To our surprise, every single time we measured the information going from premotor to primary motor cortex, we saw just as much information going from primary to premotor cortex. This went completely against the hypothesis of hierarchy and told us that the interactions between these brain regions was much more bidirectional than previously thought.

So what we did next was build a recurrent neural network model. We wanted to see if we could train it to produce both of the seemingly contrasting results we observed in our real data, the assymetric influence with perturbations and perfectly-symmetric functional connectivities. It worked, and when we went to look inside the model, we found that our simulated premotor cortex had stronger within-region inhibitory connections amongst its neurons. This finding allowed us to conclude that asymmetric reciprocal influence can emerge in these networks from local recurrent connectivity, rather than asymmetric interregional connectivity, providing a novel framework for how neural populations organize at the mesoscale level.

check out the paper

close

The retina does an immense amount of processing of visual information before it even sends signals to the rest of the brain. There are many types of retinal ganglion cells (the cells that then project to other areas in the brain) and they all are specialized to extract different features from a visual scene as quickly and efficiently as possible.

Here, we study a few different types of RGCs and learn how the spiking patterns that are specific to them help optimize the type of information they extract. Bursty suppressed-by-contrast cells (bSbCs) will burst at a baseline rate and actually decrease their activity when presented with a stimulus. This requires the downstream neurons to detect when a gap in spikes occurs and how to detect that gap quickly and reliably. It turns out that the particular way bSbCs organize their spikes allows gaps in their activity to be detect with the highest fidelity.

check out the paper

close

Coming eventually! But maybe not! Idk if what I'm trying is going to work. close

Coming hopefully actually soon ish close