An Abstraction Metric for Language Model Features
Developed and evaluated a novel abstraction metric for sparse auto-encoder features, streamlining the identification of conceptually deep and relevant features in large language models. Conducted experiments on Gemma-2-2B.
CLIP2CLAP: Joint Embedding Space Alignment for Image to Audio Generation
How similar are two joint embedding spaces? Can a small feed forward network map one to the other?
SkateMAE: Synthetic Data for Skateboard Pose Estimation
Skateboard pose estimation is a novel problem but lacks any labeled data. Is synthetic data the solution?
Can combining the techniques of two different state of the art MRI registration techniques shave a few points off the evals? Not this time.