Research

The core principle of my scientific practice is best encapsulated by Feynman's famous quote,“What I cannot create, I do not understand.” This is why I study the brain with in-silico models. With them I can simulate the brain, then cut some connections and poke some brain regions, none of which is possible in the human brain. Once I obtain a circuit model of the brain with sufficient alignment to the biological data, I can also make it a recurrent neural network (RNN) that solves tasks, to see how the brain actually computes.

My aim goes beyond understanding the brain. We should then be able to know how to fix it when it goes wrong. Psychopathology is a major societal problem and is becoming more so. Once we have a faithful in-silico brain model, we should be able to understand the pathologies in more precise terms: what circuits, dynamical properties, and regions underlie schizophrenia, bipolar disorder, and autism.

This is a long journey, and its first chapter is already written. The hierarchy that I retrieved from fMRI in the human brain ordered brain regions from the sensory and motor regions (the most external) to the paralimbic regions (the most internal), which sit right above the limbic system and the subcortex. This implies that the most important signals come from the regions that process internal, bodily information. It is empirical evidence for a theoretical perspective (which I co-led in an incredible team), grounded in cybernetics, on how to build artificial intelligence in a life-like way, so that it behaves like us, with shared directions. There we proposed that internal signals form the universal context for cognition (more below).

State-dependent signal flow hierarchy in the human cerebral cortex

The signal-flow hierarchy that I recovered from fMRI aligned significantly with the cortical hierarchy proposed by M. Mesulam in his textbook Principles of Behavioral and Cognitive Neurology, which was used in the seminal paper by Margulies et al. (2016) to define the “functional hierarchy.” However, that was an implicit hierarchy with no defined direction, and it was ambiguous about the role of the paralimbic cortex. Furthermore, my signal-flow hierarchy showed the quintessential nature of flexibility in the brain by demonstrating state-dependent changes. Importantly, the overall structure of the hierarchy (the ordering from the most external to the most internal) did not change, indicating the universal role of internally driven top-down modulation of external signals. If you are interested in learning more, click below for the full text.

Nature Neuroscience, 2026

Full text (free)Publisher

Life-inspired interoceptive artificial intelligence for autonomous and adaptive agents

This was the product of incredible teamwork. To develop this theoretical account, covering fields from cybernetics, computational biology, and machine learning to active inference, we held a weekly discussion club from 2021. During this, we invited Karl Friston (a co-author of this paper) to share his ideas with us, and he gave invaluable input throughout. The core idea of this perspective is that, for autonomous and adaptive agents, we should build the principle of survival, in a mathematically abstracted form, into artificial agents. The benchmark paper for this perspective is out at NeurIPS 2025 (Datasets and Benchmarks track). Click below for the full text, if you want to gain more insight.

Nature Machine Intelligence, 2026

Full text (free)Publisher Benchmark paper (NeurIPS 2025)