This a small first step in my PhD, but I am glad to bring something to the scientific community that might make neural circuit models a bit more reproducible, efficient and stable.

And with our refined models... we will get hands-on with cognitive stuff soon😉

Stay tuned!

We show an example of this effect by reproducing a simulation study already published. Found nets can lose up to 95% of connections (the image above was an example) and impair training!

And we also show a good simple alternative (check the preprint!😉)

A 'perfect storm' unleashed upon my simulations. Our RNNs set excitatory-inhibitory nodes by applying ReLU on weights... and leads to "dying ReLU weights"💀. Lit like pruning a tree.

So nets can break in some cognitive tasks with high energy demand (e.g., many memory items)...

Imagine you build an RNN with all node-to-node connections that mimicks the brain with excitatory-inhibitory nodes. Then, you plot the weight matrix and see this... All purple dots are weights = zero. My network barely has any excitatory (EXC) connections left after training?🥲

A new trend in neurocomputational models is adapting the power of machine learning tools to mimick the brain: you trade a bit of realism for the flexibility to train networks in many tasks 🚀

But with great tech comes great responsibility.

From my psych + methodology/stats background wanted to hop into computational cogsci and study memory and consciousness in neural circuits.

But I found my simulated networks were... dying?😅