---

Overview

in this lab we build models of the brain (often with spiking neurons), and we conduct experiments on those models to create theories of how the brain computes.

so what are spikes? also known as action potentials, spikes are rapid electrical signals that neurons generate to communicate with one another. they’re called spikes because that’s the shape of the plot (specifically the plot of the neuron’s membrane voltage over time).

currently, we are studying how multiple types of neurons contribute differently to computations in our models, and how different ways of training the models leads to different end behaviors.

Publications

Qu et al. 2026 compares four different ways of training neural network models, from most to least biologically realistic. our major finding is that more realistic training methods lead to higher dimensionality in trained models, which has some serious implications for using trained models as proxies of real brains.

zhu et al. 2026 is about how changes in inhibitory connectivity are crucial for spiking neural network models to learn even a simple task.

bojanek & zhu et al. 2020 is about how three-neuron (triangle) patterns support stable activity in spiking neural network models.

---

Presentations

cns 2026, halifax: ivyer presented on The Role of Training Paradigms in Shaping Connectivity and dynamics in RNNs. her work compares gradient-based, evolutionary, and local learning methods (check out the associated preprint). We won the Outstanding Poster award for our poster!

neuroMONSTER 2026, rome: dashiell presented updates on the role of sleep-like oscillations in RNN task learning, which he first presented as a poster at deepmath 2025, Ann Arbor on nov 7, 2025.

society for neuroscience 2025, san diego: danny presented an early career poster on the potential use of dynamic fMRI to characterize parkinson’s disease on nov 15, 2025.

6th international conference on the mathematics of neuroscience and AI, split, croatia: ezra & yotam presented their novel implementation of attention in dnns, combining both top-down and bottom-up mechanisms.

American physical society march meeting, Anaheim: ulas presented a poster on the computational benefits of multiple inhibitory neuron types on mar 19, 2025.

Bernstein conference 2024, frankfurt: Ulas and ivyer each presented posters (here and here) on Sep 30 and oct 1, 2024.

Society for Neuroscience 2024, chicago: dora and antara presented together on the relative changes in recurrent vs. output connectivity when using different learning methods in rnns at the early career poster session on oct 5, 2024.

---

People

yuqing zhu, phd

Daniel Yoon

dashiell fairborn

joyce chen

cleo yan

riley yoo

eric grossman-glover

alan jiang

hanna shigemitsu

Alumni

ulas ayyilmaz (ms in computational neuroscience at BCCN Berlin)

antara krishnan (bs in Physics at Harvey Mudd)

jiayue dora li (MS in Applied Data Science at UChicago)

yotam twersky (swe at Waymo)

ezra ford (swe at Microsoft)

perri mcelvain (PhD in neuroscience at USC)

ivyer qu (dental school next!)

julianne louie (swe at bloomberg)