We compare representations in biological and artificial visual systems to understand what makes them similar, where they differ, and which biological constraints matter for learning and behavior.
Galen Pogoncheff
PhD CandidateCS
Galen Pogoncheff is a PhD candidate in Computer Science at UC Santa Barbara. He studies how the internal representations of biological and artificial neural networks give rise to behavior.
His research spans NeuroAI, representation learning, and AI alignment. He has developed biologically constrained neural networks to identify computational principles that make artificial models more brain-like, and his current work asks how the structure of learned representations can be used to transfer desirable behaviors such as robustness between models. His broader goal is to uncover general principles linking representation, computation, and behavior in intelligent systems.
Galen earned his B.S. and M.S. in Computer Science from the University of Colorado Boulder. He has completed multiple research internships at Meta, including work on machine learning for neural interfaces, and is a recipient of the UC Santa Barbara CCS Community Fellowship and the Computer Science Outstanding TA Award.
Outside of the lab, Galen enjoys spending time in the mountains and at the gym.
Education
- PhD in Computer Science UC Santa Barbara2022–2027 (expected)
- MS in Computer Science University of Colorado, Boulder2020
- BS in Computer Science University of Colorado, Boulder2018
Honors & Awards
- Research Fellowship MATS Program (2026)
- Community Fellowship College of Creative Studies, UC Santa Barbara (2026)
- Summer Internship Meta (2024, 2026)
- Outstanding TA Award Computer Science, UC Santa Barbara (2024)
Research Areas
Affiliated Research Areas
Closed-Loop Control of Visual Neuroprostheses
We develop methods for choosing stimulation patterns that produce a desired percept or neural response, using predictive models, optimization, and feedback from the user or the brain.
Publications
Percept-aware surgical planning for visual cortical prostheses with vascular avoidance
We present a percept-aware framework for surgical planning of cortical visual prostheses that formulates electrode placement as a constrained optimization problem in anatomical space.
Galen Pogoncheff 
Alvin Wang 
Jacob Granley 
Michael Beyeler
Medical Image Computing and Computer Assisted Intervention (MICCAI) ‘26
Predictor construction can reverse multimodal neural contrasts
We show that apparent differences in what neural activity represents can reverse when language inputs are better matched, meaning that model design can strongly shape representational conclusions.
Lucas Nadolskis Galen Pogoncheff Michael Beyeler
arXiv:2609.16430
BIRD: Behavior induction via representation-structure distillation
We introduce BIRD (Behavior Induction via Representation-structure Distillation), a flexible framework for transferring aligned behavior by matching the internal representation structure of a student model to that of a teacher.
Galen Pogoncheff Michael Beyeler
The 14th International Conference on Learning Representations (ICLR ‘26)
Beyond sight: Probing alignment between image models and blind V1 Spotlight Talk
We present a series of analyses on the shared representations between evoked neural activity in the primary visual cortex of a blind human with an intracortical visual prosthesis, and latent visual representations computed in deep neural networks.
Jacob Granley Galen Pogoncheff Alfonso Rodil Leili Soo 
Lily M. Turkstra Lucas Nadolskis Arantxa Alfaro Saez Cristina Soto Sanchez 
Eduardo Fernández Michael Beyeler
Workshop on Representational Alignment (Re-Align), ICLR ‘24
Explainable machine learning predictions of perceptual sensitivity for retinal prostheses
We present explainable artificial intelligence (XAI) models fit on a large longitudinal dataset that can predict perceptual thresholds on individual Argus II electrodes over time.
Galen Pogoncheff 
Zuying Hu 
Ariel Rokem Michael Beyeler
Journal of Neural Engineering
Explaining V1 properties with a biologically constrained deep learning architecture
We systematically incorporated neuroscience-derived architectural components into CNNs to identify a set of mechanisms and architectures that comprehensively explain neural activity in V1.
Galen Pogoncheff Jacob Granley Michael Beyeler
37th Conference on Neural Information Processing Systems (NeurIPS) ‘23
