Sequential auto encoding of neural embeddings - SANE

• learn compressed latent representation of model sequence
• contrastive loss
• capacity scales with model size
• gracefully scale to larger models with different architectures
• different models share embedding space
• unified for generative and discriminative