PASTA is an innovative AI training system designed to adapt to individual user preferences through a two-stage strategy. Gathering vast interaction data is essential but challenging due to privacy concerns. To overcome this, PASTA combines real human feedback with user simulation. Initially, it utilizes over 7,000 rated interactions, including prompt expansions from the Gemini Flash model and images from the Stable Diffusion XL, creating a foundational dataset. This data trains a user simulator that mimics real preferences and choices.
At the core of PASTA is a user model, consisting of a utility model and a choice model, each predicting user preferences for images. By employing pre-trained CLIP encoders and an expectation-maximization algorithm, PASTA identifies unique user types with similar tastes, generating over 30,000 simulated interaction trajectories. This comprehensive approach allows PASTA not only to gather more data but also to explore diverse user behaviors, enhancing its collaborative capabilities with users.
Source link
