NeurIPS oral: Neural population tuning statistics as priors for multitask generalization
We are very happy to share that Zach Cohen’s paper Neural population tuning statistics as priors for multitask generalization has been accepted as an oral presentation at NeurIPS 2026. Congratulations, Zach!
Neurons in a population differ widely in their tuning: in their selectivity, their gain, and in how their preferred stimuli are spread across stimulus space. This heterogeneity has mostly been studied for its impact on how precisely a population encodes information. The paper instead asks how it shapes learning. It develops a probabilistic theory showing that, in large populations, the tuning statistics act as a prior over functions that predicts how well the population can learn and generalize across a family of related tasks. Building on this, it provides a Bayesian account of how tuning should adapt during learning, which explains why hippocampal place cells come to over-represent rewarded locations with experience.
Zach will present the work in a talk on Tuesday, December 8 (4:00–4:15pm AEDT, Pyrmont Theatre) at NeurIPS in Sydney, followed by a poster that evening in Poster Session 2 (5–8pm, Hall 1-4). For short summaries, see Zach’s Bluesky thread and Jan’s thread.