User Friendly and Adaptable Discriminative AI: Using the Lessons from the Success of LLMs and Image Generation Models

Discriminative methods focus on modeling the conditional probability of outcome(s) given a context (such as a feature vector). In contrast, generative methods focus on modeling the joint distribution of data. Discriminative models have historically found success in classification and regression tasks in various domains (e.g., finance, healthcare, automotive, etc). On the other hand, newer generative models, such as Large Language Models (LLMs) and diffusion models, have succeeded in open-ended tasks that require versatility and creativity in addition to traditional prediction tasks. We hypothesize that the value of these new generative models is enhanced because they are user-friendly and highly adaptable, making it easier for non-experts to interact with them and produce valuable results with minimal effort. However, this is not the case with current discriminative models. In this work, we explore ways to make discriminative models more user-friendly and adaptable, which we hypothesize will increase their adoption in more applications and bring them on par with the success levels seen with generative AI tools.