Cartesia· Sonic / InkResearch paper

Cartesia announced a research collaboration on hierarchical networks (H-Nets), a new architecture that learns to segment and compress raw data into meaningful concepts. H-Nets scale better with data than Transformers with BPE tokenization, are more robust to input perturbations, and can be stacked for deeper hierarchies. The company released checkpoints for H-Net models on HuggingFace.

From the source

We’re excited to announce our latest research collaboration on hierarchical networks (H-Nets) , a new architecture that natively models hierarchy from raw data.

cartesia.ai