Chizat, L. & Bach, F. (2018). Advances in Neural Information Processing Systems 31 (NeurIPS 2018). arXiv:1805.09545.
Read on 2026-10-29 (NOTE-1367): the paper in full, the supplement in part.
Why this is here
It is also held in the Anthology of the SOTA, as ANTH-LIT-298. Holding it in both follows ADR-013. The anthology reads it for what it says about training over-parameterized models. This record reads it for a different question: whether a population of learned features, treated as particles whose density evolves under a gradient flow, can be analysed as an ecological community whose members differentiate and persist or are excluded. That question belongs to QUESTION-132 and QUESTION-135, and CLAIM-369 is the finite, symmetric case. The synthesis the owner relayed on 2026-10-18 recommended it as “the learning-theoretic side of the adaptive population model”. A reading should check whether the mean-field flow’s stationary points satisfy an invasion-type condition, a new unit entering where the first variation of the objective is negative, as the Lotka–Volterra potential case does.