John Hershey

About

Key figure spanning classical and deep eras of source separation. Contributions: deep clustering (2016), permutation invariant training, MixIT (mixture invariant training). Google Research.

Relevance

Foundational contributions to deep source separation. PIT is a core training technique used in Demucs and other modern separation models. MixIT enables training without isolated stems — potentially valuable if isolated bluegrass stems are scarce.

Mentions

• ../concepts/deep-clustering-separation — 2016
• ../concepts/permutation-invariant-training — core training technique
• ../concepts/mixture-invariant-training — trains without isolated stems