Welcome! I'm Pavel Tolmachev

Postdoctoral Research Associate at Princeton Neuroscience Institute

I build models of neural systems and take them apart again. Most of my work is on recurrent neural networks: training them on cognitive and motor tasks, then reverse-engineering the circuits they converge on. My recent 2025 Nature Machine Intelligence paper showed that something as innocuous as the choice of single-unit nonlinearity sets the computational primitives a network builds with, which in turn results in different circuit-solutions. Although networks implementing those different circuit-solutions matched on in-distribution performance, their behavior diverges meaningfully once you probe them with previously unseen stimuli. So if we want insight about the brain, we should be using models aligned with biology.

Before my postdoc, I spent a PhD in electrical engineering modeling brainstem central pattern generators — the half-centre oscillators that keep breathing and swallowing from mis-coordination. The through-line is the same: fit dynamical systems to messy biology, then interrogate the model until it gives up testable mechanistic predictions. I wrote and maintain trainRNNbrain, an open-source package for training and dissecting RNNs trained on cognitive tasks.

This trajectory has naturally taken me toward digital twins of neural systems: models faithful enough that you can perturb them in silico and have the resulting activity match what the real system would have done. I'm further interested in building hierarchical mechanistic models of those systems — because a model only replaces one system with another, albeit one more amenable to analysis. How do we then understand it?

Away from the papers, my interests are diverse. Electronic music and drum and bass. Meditation, the philosophy of mind, and conscious experience — perpetually fascinated and profoundly puzzled by the Hard Problem. Metabolic health and longevity (did a whole genome sequencing and then obsessively mined it). And crypto: a long-standing Bitcoin supporter, and taken with Ethereum's ambition of building a world state-machine.