Neuro-vector-symbolic vision transformers

Hyperdimensional computing for few-shot classification and generation

Hyperdimensional computing grafted onto a vision transformer backbone, adding on-chip learning capability while extending HD representations to generative settings.

  • A ladder-side HD computation path attached to a standard transformer backbone, giving the model few-shot learning capability.
  • HRR-based vector-symbolic algebra used to retrieve input data generatively, avoiding the cost of reading out prototypes from distributed memory.
  • Design-space exploration across 2D, 2.5D and 3D integration for accelerating the model with separate FFN, attention and HD cores.