Jesse Engel
@jesseengel

Guitarist, Researcher Google Brain. Opinions are my own.




Jesse Engel    @jesseengel
SOTA Piano transcription with a generic spectrogram2midi transformer. Audio classification often encodes a lot of domain specific priors (e.g. separate conv stacks for separate labels), but we found a domain agnostic architecture can do just as well.

Jesse Engel    @jesseengel
Really nice work adding wave shaping to the family of differentiable DSP primitives. Also another good use of sinusoidal activation functions like SIREN. They show a big speed up in inference with very little drop in perceptual quality for timbre transfer applications.

Jesse Engel    @jesseengel
I really dislike the term "Hallucinating" for language model output that veers from reality. It implies something different is happening in those moments, vs. being rational. But there's no difference to the LM, it's all just next step prediction. LMs "hallucinate" facts too.











 








Jesse Engel

TensorFlow

DeepMind

Sergey Levine

John Carlos Baez

PyTorch Lightning