ProbXiv
Explorersign in
Problem archiveProblem record

Statement

Fix the depth LL, the endpoint widths d0d_0 and dLd_L, and the activation degree rr. Let d53d=(d0,d1,d55fd53dots,dL)d53d=(d_0,d_1,d55fd53dots,d_L) be a width vector, and let d53d\preccurlyeqd53d′d53d\preccurlyeqd53d' mean coordinatewise comparison. An architecture is filling when its neurovariety d4b1d53d,rd4b1_{d53d,r} has the full expected ambient dimension; d53dd53d is minimal when no smaller width vector with the same fixed endpoints has this property.

Unimodality conjecture. Every minimal width vector whose architecture is filling is unimodal: there exists i∈{0,1,d55fd53dots,L}i\in\{0,1,d55fd53dots,L\} such that (d0,d55fd53dots,di)(d_0,d55fd53dots,d_i) is weakly increasing and (di,d55fd53dots,dL)(d_i,d55fd53dots,d_L) is weakly decreasing.

The conjecture formalizes the architectural heuristic that a network should first expand its feature capacity and then contract it. The supplied text gives motivation but no proof or resolution.

Record

Comments

Nothing has been published against this problem yet, and nothing has been said about it.

Sign in with an institutional address to take part in the discussion. Reading every thread stays open to everyone.

Sign in

Solve with an agent

Open the statement in a chat, with the problem and the ground rules already written into the prompt.

This opens a third-party site. Nothing is posted back to ProbXiv and nothing you write there is recorded here — what a model gives you is an attempt, which a person still has to check.