Schmidhuber renews residual connections credit claim: "LSTM is a ResNet rotated 90 degrees"

SchmidhuberAI · x · 2026-10-15

Schmidhuber shared a video quoting Ilya Sutskever: "an LSTM is a ResNet rotated 90 degrees." He uses it to restate his timeline for residual connections: in 1991, Hochreiter solved the vanishing gradient problem via recurrent residual connections (weight 1.0); the 1997 LSTM used plain recurrent residual connections; a 1999 variant (with Felix Gers) introduced gated recurrent residual connections. This echoes his long-standing claim that the century's most-cited paper, deep residual learning, traces back to 1990s LSTM work.

Related event: Yang Quotes Ilya's 'LSTM Is a Rotated ResNet'; Schmidhuber Cites History(2 posts)→

Original post →

More from Fun

Fun channel →