Learn to Compress or Compress to Learn? Shannon and Turing meet again in the LLM era

Abstract

The year 1948 gave birth to two transformative pillars of modern computing: Claude Shannon's information theory, which gave rise to data compression, and Alan Turing's writings on intelligent machinery, which ignited the field of artificial intelligence. Today, nearly eight decades later, we are witnessing a profound hybridization of these once-distinct fields. As we push to efficiently store, process, and retrieve massive datasets across diverse data types, users, devices, and systems — from vector and graph databases to cutting-edge agentic memories — AI and compression can no longer operate in silos. In this talk, we investigate how Learning is fundamentally reshaping the fields of Indexing and Compression. We will focus on two challenging research directions: designing Learned Data Structures that fuse classical indexing with machine learning to achieve unprecedented efficiency, and developing next-generation Data Compressors driven by powerful autoregressive and diffusion neural models to reach new frontiers of efficacy. We will also examine the reverse direction — how Compression can reshape Learning through model compaction and comprehension. By unpacking the distinct computational advantages and current practical pitfalls of these lines of investigation, we will outline a roadmap for these research frontiers and hopefully inspire new cross-disciplinary collaborations.