Necessary-and-sufficient closedness conjecture for shallow sparse ReLU networks

Let Ω=[B,B]N0\Omega=[-B,B]^{N_0} with B>0B>0, and let I=(I2,I1)\boldsymbol{I}=(I_2,I_1) be a shallow-network architecture. Let FI(Ω)\mathcal{F}_{\boldsymbol{I}}(\Omega) be its function space. The function space is considered with the uniform norm in

(C0(Ω),).(C^0(\Omega),\|\cdot\|_\infty).

Shallow-network closedness conjecture. The set FI(Ω)\mathcal{F}_{\boldsymbol{I}}(\Omega) is closed in (C0(Ω),)(C^0(\Omega),\|\cdot\|_\infty) if and only if I\boldsymbol{I} satisfies the second condition of the stated generalized closedness theorem.

This conjecture seeks a necessary and sufficient condition for closedness in the shallow case L=2L=2, where the paper has separately obtained necessary and sufficient conditions. The source does not state a resolution.

Sources & referencesView supporting material

Primary source

Quoc-Tung Le, Elisa Riccietti and Rémi Gribonval, “Does a sparse ReLU network training problem always admit an optimum?”, arXiv:2306.02666 (2023).

Progress summary

Never refreshed

Nothing recorded yet. Refresh searches the literature and the public web for attempts on this problem, and writes the first summary here.

Solutions 0

No solutions have been posted yet.