This project distills a model's word embeddings into human-interpretable "concept-vectors", i.e. vectors in which each component tracks concerns like semantics, syntax, and even statistics potentially, while associating each component with a human readable and human definable label.
See the readme[1] for details.
I want to acknowledge explicitly that this is a data design project. I have quite a bit of experience with data transformation and manipulation, but limited experience with NNs. I have not tested this on models, and I currently don't have the resources to build a comprehensive database to test it on models. I'm posting primarily for human feedback/criticism, and simply to share the idea since this is as far as I can currently take it.
Hi HN,
This project distills a model's word embeddings into human-interpretable "concept-vectors", i.e. vectors in which each component tracks concerns like semantics, syntax, and even statistics potentially, while associating each component with a human readable and human definable label.
See the readme[1] for details.
I want to acknowledge explicitly that this is a data design project. I have quite a bit of experience with data transformation and manipulation, but limited experience with NNs. I have not tested this on models, and I currently don't have the resources to build a comprehensive database to test it on models. I'm posting primarily for human feedback/criticism, and simply to share the idea since this is as far as I can currently take it.
[1] https://github.com/truehumanexe/concept_vector/tree/main