\(\text{(a) PerceiverIO}\)
\(\text{(c) LNO}\)
\(\text{(b) Transolver}\)
\(\text{(d) FLARE}\)
\((B\text{ blocks})\)
\(\text{Encoder } (N \to M)/ \text{ decoder } (M\to N)\)
\(\text{Feature}\\\text{projector}\)
\(\text{Cross attn}\\\text{projector}\)
\(\text{Input/output}\)
\(\text{(\textit{N} tokens)}\)
\(\text{Token sequences}\)
\(\text{Self attention}\\(M \to M)\)
\(\text{Latent}\)
\(\text{(\textit{M} tokens)}\)
\(\text{Cross attention}\\(N \to M)\)
\(\text{Attention Operators}\)
By Vedant Puri
PhD student at Carnegie Mellon University
