willbaskett

ChemROAR. A novel generative embedding architecture. Find clusters of molecules with specific properties and then generate new molecules from those clusters.

MedROAR. A multi-modal embedding model for hierarchically clustering medical records. Identifies a patient's clinical subgroup using whatever records are available at the time.

grokking_in_both_directions. Data is valuable and finite. How much performance is lost when you only train right-to-left in your autoregressive models?