Deterministic Learner

Deterministic learners, in contrast to their randomized counterparts, make predictions based solely on the input data without incorporating randomness. Current research focuses on understanding their performance limitations, particularly in adversarial settings like online classification and data poisoning, and exploring the trade-offs between deterministic and randomized approaches in various learning paradigms, including bandit feedback and multiclass scenarios. Key areas of investigation involve characterizing optimal mistake bounds using combinatorial measures like the Littlestone dimension and developing novel deterministic algorithms, such as those based on invertible layers in generative models, to improve performance and address challenges posed by strategic agents or adversarial manipulations. These advancements contribute to a deeper understanding of fundamental learning limits and inform the design of robust and efficient learning systems.