cv
General Information
| Name | Ann Huang |
| Location | Cambridge, MA, USA |
| annhuang@g.harvard.edu |
Education
-
2023.08 – Present PhD in Neuroscience
Harvard University - Supervised by Kanaka Rajan
-
2019.09 – 2023.05 Bachelor of Science
McGill University - Honours Neuroscience; Minor in Mathematics and Computer Science
Research Experience
-
2024.03 – Present Graduate Researcher, Rajan Lab
Harvard, Cambridge, MA, USA - Supervised by Kanaka Rajan
- Developed a systematic framework for measuring and controlling solution degeneracy in task-trained RNNs, identifying four factors that mediate inter-seed variability at the level of behavior, neural dynamics, and weights, which serves as a principled guideline for when a mechanism found in one model transfers to another (NeurIPS 2025 Spotlight)
- Built InputDSA, a similarity metric for separating and comparing intrinsic vs. externally driven dynamics in complex systems, applicable to both neural data and model activations (ICLR 2026)
- Currently developing Hessian Null-Space Continuation, a scalable method that traverses the space of functionally equivalent networks and steers models toward mechanisms with desired internal properties at fixed behavior
-
2021.09 – 2024.06 Research Assistant, Richards Lab
McGill & Mila, Montreal, QC, Canada - Supervised by Blake Richards
- Developed Stochastic Genomic Bottleneck, a generative method that compresses task-relevant priors into RNN weights, enabling rapid learning from minimal data — a structured-prior analog of pretraining (bioRxiv)
- Interpreted the memory mechanisms and navigation strategies implemented by deep RL agents via decoding, dynamical systems analysis, dimensionality reduction, and ablation (Scientific Reports)
-
2022.06 – 2023.05 Visiting Researcher, Mathis Lab
EPFL, Geneva, Switzerland - Supervised by Alexander Mathis
- Designed a goal-conditioned hierarchical deep RL algorithm for musculoskeletal control, combining a pretrained language model that proposes behavioral subgoals with a distributed actor-critic architecture that executes the low-level actions
- Contributed to latent exploration methods that shape covariance across actuators for coordinated exploration and control (NeurIPS 2023)
-
2021.09 – 2022.06 Research Assistant, Gershman Lab
Harvard, Cambridge, MA, USA - Supervised by Sam Gershman
- Developed a theory that explains action chunking as policy compression under the rate-distortion framework; designed cost-sensitive actor-critic RL models with information-theoretic penalties on policy complexity
- Validated models against human behavior via Bayesian model selection and simulation (Cognition)
-
2019.11 – 2021.05 Research Assistant, Brandon Lab
McGill, Montreal, QC, Canada - Supervised by Mark Brandon
- Studied long-term stability of spatial context representations in the hippocampus despite neural drift
- Built a Bayesian decoder for animal location from neural recordings in mice; used representational similarity analysis and generalized linear models to factor the neural representations of animal location into orthogonal contextual and temporal-drift components