Andrew Mah

GitHub | Google Scholar | LinkedIn | Email

prof_pic.jpg

Center for Computational Neuroscience

Flatiron Institute

New York, NY 10010

About Me

I’m a Flatiron Research Fellow at the Flatiron Institute Center for Computational Neuroscience, working with Alex Williams. I study how neural networks represent and use information to perform inference, using small, but surprisingly powerful, models that I can analyze in closed form.

Two recent results:

  • Why a linear state-space model fails to length-generalize in a hierarchical language task. Across seeds, linear RNNs + MLP readout trained on bounded-depth Dyck languages varied widely in how well they extrapolated. I traced the failures to a slow drift mode in the linear recurrence that gradually pushes the representation out of the region where the readout is accurate. Clamping that mode at inference time, with no retraining, restores performance at up to 100x the training length.

  • When linear models are enough. We proved error bounds showing that forward filters on a broad class of hidden Markov models can be well approximted by linear recurrent models, and validated the bounds on 230 randomly generated and adversarial HMMs (NeurIPS 2026). The bounds depend on easily computable properties of the HMM, which gives a practical check for when a task actually requires nonlinear inference.

These questions are inspired by my Ph.D. work at NYU with Christine Constantinople, where I studied how uncertainty about hidden states shapes learning and decision-making in the brain. I now ask what artificial networks infer about hidden structure and how they use it. So even though I’ve traded rats for JAX and behavioral rigs for GPU clusters, the underlying questions remain the same.

I’m seeking research scientist roles in interpretable AI, sequence models, and probabilistic reasoning.

Selected Publications

Mah, A.*, Pughe-Sanford, J.L.*, Harvey, S.E., & Williams, A. (2026) Linear approximations to HMM filtering Advances in Neural Information Processing Systems.

 

Mah, A., Golden, C.E., & Constantinople, C.M. (2024). Dopamine transients encode reward prediction errors independent of learning rates. Cell Reports, 43(10). DOI: 10.1016/j.celrep.2024.114840

 

Mah, A., Schiereck, S.S., Bossio, V., & Constantinople, C.M. (2023). Distinct value computations support rapid sequential decisions. Nature Communications, 14(1), 7573. DOI: 10.1038/s41467-023-43250-x