Neural computation
How do large populations of cortical neurons implement memory, decisions, and behavior? I develop low-dimensional theories and data-constrained recurrent models that can be compared with recordings and perturbations.
Fatih Dinc
Joint Postdoctoral Scholar at the Kavli Institute for Theoretical Physics and the Geometric Intelligence Lab, UC Santa Barbara
I study how biological and artificial neural networks implement computation. My work combines dynamical systems, optimization, and large-scale neural data to make learning and memory mechanisms interpretable.
About
I work at the intersection of neuroscience, theoretical physics, and computer science. I combine mathematical theory, machine learning, and brain-imaging data to understand how distributed neural activity produces learning and memory. I use the same principles to make artificial neural networks more interpretable.
At UC Santa Barbara, I work with Boris Shraiman on modeling morphogenesis using AI, focusing on how local interactions among cells give rise to biological form. With Nina Miolane, I study how large, high-dimensional populations of neurons carry out much smaller, behaviorally meaningful computations. We develop theory and data-analysis methods to identify low-dimensional variables that can be read out from population activity, characterize the dynamics that update those variables, and determine when very different neural networks implement the same underlying computation.
During my PhD in Applied Physics at Stanford, I worked with Mark Schnitzer and experimental collaborators to study memory formation and retrieval using large-scale brain-imaging data from mice. I developed tools for extracting neural activity and fitting data-constrained recurrent networks, including EXTRACT and CORNN, and used dynamical-systems models to study learning and short-term memory.
At NTT Research, I worked with Hidenori Tanaka on foundational problems in interpretable AI and optimization, including why neural networks can learn abruptly and how their internal dynamics change during training. My earlier training at Perimeter Institute, the University of Waterloo, and Boğaziçi University was in theoretical physics and electrical engineering.
- 2025–present
- UC Santa Barbara · Postdoctoral Scholar
- 2019–2025
- Stanford University · PhD, Applied Physics
- 2018–2019
- Perimeter Institute / Waterloo · MSc, Physics
- 2014–2018
- Boğaziçi University · BS, Physics and Electrical Engineering
Highlights
- 2026 SIAM Richard C. DiPrima PrizeOutstanding doctoral dissertation in applied mathematics · UCSB News profile
- Developer of EXTRACT - since 2021Open-source neural data analysis software used by hundreds of laboratories
Research
Mechanistic interpretability
How do neural networks learn and represent algorithms? I study attractor geometry, abrupt learning, memory mechanisms, stability, and trainability in biological and artificial networks.
Selected work
Recent work in theoretical neuroscience, mechanistic interpretability, machine learning, and neural data analysis.
2026
A geometric and dynamical theory of latent computations in biological neural networks
bioRxiv
2026
Low-dimensional neural codes suppress neuronal noise and extend the working memory duration
bioRxiv
2026
Predictive pursuit emerges in high-dimensional recurrent neural networks
bioRxiv
2026
Multi-modal alignments of in vivo imaging and spatial biology datasets at cellular resolution
bioRxiv
2026
A ghost mechanism: An analytical model of abrupt learning in recurrent networks
Physical Review X
2026
Amantadine modulates action-specific neural ensembles in hypokinetic and hyperkinetic conditions
Movement Disorders
In review
Distinct spiking sequences mediate global brain state transitions
Under review
In review
Synaptic Strength Controls Trainability and Structural Stability in Rank-Deficient RNNs
Under review
2025
Extracting task-relevant preserved dynamics from contrastive aligned neural recordings (CANDY)
NeurIPS · Spotlight
2025
Dynamical phases of short-term memory mechanisms in RNNs
ICML
2025
Understanding and controlling the geometry of memory organization in RNNs
arXiv
2025
Beyond Euclid: An illustrated guide to modern machine learning with geometric, topological, and algebraic structures
Machine Learning: Science and Technology
2024
EXTRACT: Fast, scalable, and statistically robust cell extraction from large-scale neural calcium imaging datasets
bioRxiv
Open-source software now used by hundreds of laboratories.
2024
ActSort: An active-learning accelerated cell sorting algorithm for large-scale calcium imaging datasets
NeurIPS
2024
Neural circuit basis of placebo pain relief
Nature
2024
Global Distortions from Local Rewards: Neural Coding Strategies in Path-Integrating Neural Systems
NeurIPS
2023
CORNN: Convex optimization of recurrent neural networks for rapid inference of neural dynamics
NeurIPS