Minima of linear networks have 2^{l-1} connected components, skip connections can reduce this count, and rescaling symmetries can make linear interpolation between minima fail.
By Lemma 5.1, any g ∈ GL(h) with det(g) < 0 can bring (W1, W2) and (W ′ 1, W′
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Understanding Mode Connectivity via Parameter Space Symmetry
Minima of linear networks have 2^{l-1} connected components, skip connections can reduce this count, and rescaling symmetries can make linear interpolation between minima fail.