Linear Mode Connectivity Sparks Debate Over Its Origins
A Google researcher called linear mode connectivity one of the most underrated properties in neural network training, recalling a four-year-old finding that test loss stays flat along linear paths while training loss does not. Others clarified that Vaishnavh's observation predates the 2019 Frankle paper, sparking a dispute over who discovered it first.
2026-09-28 ~ 2026-09-29 · 2 related posts
- Linear mode connectivity appears on test loss but not train loss, notes researcher Arohan — _arohan_ · 2026-09-28
- Who First Saw Linear Mode Connectivity? Vaishnavh's Test-Loss Observation Predates the 2019 Paper — roydanroy · 2026-09-29