Seonglae ChoDimension Agnostic Neural Process
DANP: a Neural Process composed of DAB + Transformer + a stochastic latent path + a decoder. It takes observed (x,y) pairs and predicts the output and its uncertainty for new inputs.
Applies the same embedding and attention rules to every coordinate. Since a change in dimensionality only changes the number of tokens, unseen dimensions can be handled with the existing weights.
DAB
Dimension Aggregator Block
A block that transforms representations so that differing input and output dimensionalities can be handled. It performs per-coordinate embedding → attention → pooling over the input side. DAB itself does not contain DANP's overall meta-learning or probabilistic prediction capabilities.