Скачать или смотреть WDI25 - Adaptive machine learning models for the AI age - Patrick Deziel

In the age of generative AI, more people are interacting directly with machine learning models than ever before. This creates a few problems for machine learning engineers and data scientists. How do we build models that remain relevant in production? Can these large models be customized to meet the changing needs of individual users? How can smaller organizations train effective AI models at cost?

Real-time machine learning is a classic technique that utilizes continuous learning to train models. Unlike traditional batch models, these models learn in an ""online"" fashion by interpreting observations from their environment. Online models have the potential to keep up with shifting trends and patterns and are therefore preferable for certain use cases, such as trend forecasting, recommendation systems, and reinforcement learning with human feedback (RLHF). The success of the DeepSeek-R1 model proves that a pure reinforcement learning approach can outperform models that are trained on large, curated datasets with supervised fine-tuning (SFT). Online models can be more cost-effective because they can be trained with less compute and memory than their offline counterparts.

In this talk, I will explain how real-time machine learning models work and explore their relevant applications with respect to current industry trends. Specifically, I will show practical examples of a personalized recommendation model and a language model that learns via RLHF.