Grasping Concept Spaces
From Computer Laboratory Group Design Projects
Jump to navigationJump to search
Client: Sam Henshall-West, JAID <sam.henshall-west@jaid.io>
Today's most powerful machine learning models, including generative AI LLMs like ChatGPT, encode their knowledge as higher dimensional latent spaces. Vector-based concepts in that space can be clustered as regions on a 2D screen, for example using the t-SNE visualisation algorithm. Your task is to create a 3D version of t-SNE that is interactive, so users with VR headsets can literally “grasp” a concept, and manipulate it to explore, fine-tune, or adjust the machine learning model.