I develop neuro-AI foundation models of neural and behavioral dynamics for brain-machine interfaces toward improved clinical outcomes in behavioral disorders.

• Developing supervised ans SSL models of neural activity that generalize to future days;
• Co-leading the development of interpretable SOTA time-series forecasting models;
• Applying these models to identify biomarkers and therapeutic targets across patient groups.

Publications: 2 co-last-author (incl. NeurIPS BrainBodyFM Spotlight); 4 co-last-author in submission

Used machine-learning approaches to develop normative models of reward-driven behaviors and relate them to neuronal activity observed in the brain.

• Developed data-driven models of sequential decision-making, motivation, and conflict;
• Co-developed approaches to compressibility and self-assembly of neural network models;
• Worked on the function of olfactory receptors and the structure of olfactory connectivity.

Publications: 5 first-author (incl. 2 NeurIPS and 2 PNAS), 2 co-authored (incl. ICML)

Designed computer-vision models to identify brain-wide changes in neuronal populations during brain development and neurodegenerative disorders.

• Developed pipelines for microscopy, detection, and alignment of brain-wide cell populations;
• Used these pipelines to study how antidepressants, development, and radiation affect neurogenesis;
• The pipelines are now used in studies of traumatic brain injury and neurodegenerative disorders.

Publications: 2 first-author and 2 co-authored; led to a body of follow-up research.

Developed numerical models and worked towards measurements of electrical and thermal properties of high-current superconductive cables to pursue requirement-based design.

Patient condition in treatment-resistant OCD can be improved with deep brain stimulation. To identify the stimulation targets and biomarkers, we developed a robust automatic data processing pipeline. The processed data has allowed us to identify the disorder-related networks in the brain.

Inpatient brain recordings and reports remain sparse to minimize the burden on patients. This requires additional modeling considerations. We show that self-supervised learning of brain activity plus finetuning on behavioral data best captures the neural counterparts of behavior.

Time series are well-modeled in the wavelet domain, however, higher precision requires exponentially more features. To prevent overfitting, we propose an encoder that combines wavelet features with importance scoring. Our encoder shows SOTA results on the majority of 100+ tested datasets.

We argue that time-series datasets represent instantiations of noisy generative processes. We model these processes with diffusion and introduce the minimum-over-N metric from CV for evaluaiton. Identified nonzero uncertainty of standard benchmarks calls for their distributional modeling.

Neural activity has several distinct properties: signal structure is shared across channels, events have stereotyped shape, events are sparse. To account for these properties, we use attention, matching feature sets, and stratification, leading to biologically relevant predictions.

Time-series data can come from devices with different properties. To jointly use such data, it should be uninfied for the model training. We view this as an optimal transport problem, leading to improved domain generalization across clinically relevant biosignal benchmarks.

Chronic aggression often leads to bullying or social defeat. Why engage in such unproductive behaviors? We find in mice that individuals maintain beliefs about the strengths and beliefs of each other. Upon interactions, these beliefs shift, resulting in actions that ignore real strengths.

To make decisions, we rely on observations, experience, and motivation. What is the computational benefit of the latter, and can we use it in machine learning? We find that biological motivations speed up reinforcement learning, compress the models, and enable zero-shot learning for new tasks.

When do we stick to ongoing plans, and when do we seek better options? We show that such decisions flip when the perceived reward falls below its average over the recent past. We argue that this approach is compute-inexpensive and Bayes-optimal in ever-changing natural environments.

Odor perception is hard to model becuase it's highly nonlinear and the data is sparse. Here we use physics and chemistry to model the odor transport all the way to the nervous system. We use these approximated neural inputs in deep learning and show improvement in predicting odor percepts.

The perceived smell identity doesn't change with an increase in odor concentration despite new cells firing in the brain. How is it represented? We show that the odor identity is encoded in groups of first-responding cells and that the wiring of the fruit fly's brain supports the processing of this code.

How to predict smell knowing the molecular structure of an odorant? In human, odorants hit the smell receptors at various angles exposing various features of the molecule. We reproduced this scheme in a convolutional network to set the state-of-the-art for odor classification performance.

Hypernetworks reduce a model's parameter count while maintaining its performance. While geared at LLMs, hypernetworks have only been tested on toy tasks. Here, in theory and experiment, we show a set of considerations for that makes such scaling computationally tractable and plausible.

The brain wiring is specified by the genome, yet the information in the brain exceeds the genomic capacity. Here we show that such a gap - the genomic bottleneck - can be bridged by constraining the networks' degrees of freedom to form connectivities that are both efficient and compressible.

Connections between neurons shape the experience and the innate abilities of a person. How do these connections unfold in the brain of a newborn? Here we propose a theoretical model for such unfolding based on "barcodes" - the unique markers tagging individual neurons with distinct labels.

What's changing in the brain during disease, and how do the changes evolve? Here we introduce CORGI, an algorithm to align 3D images of the brain, "matching" brain regions and "synchronizing" brain samples. On output, it shows time-dependent maps of changes in healthy and affected brains.

Whole-brain microscopy has become a commonplace tool, generating amounts of data far beyond the throughput of expert analysis. To facilitate whole-brain studies of cell populations, we introduce DALMATIAN, an automated cell detection algorithm compatible with common types of cell markers.

Memory and cognition rely on stem cells and their division in the brain. To study stem cells across the brain, we propose using synthetic DNA "letters" that get into newly-divided cells and "click" to small fluorescent molecules, illuminating these cells brain-wide under 3D microscopes.

How to grasp the dynamics of neurogenesis? Repeated whole-brain microscopy and alignment allow reconstructing stem cell division and migration over time. We used it to find new migratory paths of stem cells in the mouse brain and describe their changes from birth to adulthood.

Antidepressants and cognition-enchansing medicine operate through altering the hippocampal neurogenesis in the brain. How do they act? We characterize the local changes in neurogenesis caused by two popular drugs, offering a stepping stone for future studies of their function.

Frequent flyers are exposed to fast neutron irradiation coming from cosmic rays in the Earth's magnetosphere. Does it put them at risk? We find that, while such irradiation does not cause memory deficit or behavior changes in mice, it suppresses the division of stem cells in their brain.