We also find that activation functions like JumpReLU and layers such as convolutional or MLPMixer offer better interpretability than Koopman autoencoders.

4/4

Sparse autoencoders improve the linearity of control vectors, which we realize with virtually no runtime cost -- just a scale-and-add.

👇 3/4

Latent-space regularities appear in motion embeddings of transformer models, and this enables vector-arithmetic operations including controlling representations.

👇 2/4

Latent-space regularities appear in motion embeddings of transformer models, and this enables vector-arithmetic operations including controlling representations.

👇 2/4