Final sharpness under gradient flow in a single-neuron deep linear network is governed by a dataset-only difficulty score Q, with bounds that also explain depth, batch size, and learning rate effects.
We denote o⋆ i =w ⊤ i u⋆ for each i∈[r]
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
ACCEPT 1representative citing papers
citing papers explorer
-
Understanding Sharpness Dynamics in NN Training with a Minimalist Example: The Effects of Dataset Difficulty, Depth, Stochasticity, and More
Final sharpness under gradient flow in a single-neuron deep linear network is governed by a dataset-only difficulty score Q, with bounds that also explain depth, batch size, and learning rate effects.