Tutorials

Presenters: Lalitha Sankar and Monica Welfert

Generative AI holds the promise of enabling unprecedented human-machine interactions through deep learning architectures trained to create synthetic data that is both relevant and meaningful. This technology is rapidly influencing modern life, driven by methodological innovations grounded in information and learning theories. Three broad generative approaches have emerged: Generative Adversarial Networks (GANs), which employ a two-agent game to implicitly learn data distributions; Diffusion models, which transform data by gradually adding noise and then learn to reverse this process, generating synthetic data with improved stability over GANs; and Large Language Models (LLMs), which use probabilistic models for next token prediction in text generation.

In this tutorial, we will take a theory-centric approach to explain the foundational principles of GANs and diffusion models, contrasting their advantages and limitations. We will provide a comprehensive overview of both methodologies, emphasizing the key mathematical formulations and theoretical challenges relevant to the information theory research community. Related methods such as variational autoencoders and flow-based models will also be briefly discussed. The tutorial will be delivered predominantly using iPad and pencil to focus on essential mathematical concepts. The presentation will also be supplemented by slides that contain results from these approaches and detailed references for further exploration.

Presenters: Joseph Boutros and Alex Sprintson

Modern cryptography was born half a century ago. Algorithms were based on the hardness of mathematical problems such as integer factorization and discrete logarithm. These foundations provided the security guarantees needed for key exchange, encryption, digital signatures, and authentication. Over time, cryptographic research has evolved to address emerging threats, leading to the development of post-quantum algorithms.

Error-correcting codes and lattices are the most famous theoretical tools employed in post-quantum cryptography. While traditional schemes are vulnerable to quantum attacks, lattice-based and code-based cryptosystems offer resilience against quantum adversaries. As quantum computing progresses, the integration of these tools drives the transition toward a quantum-safe digital infrastructure.

In this tutorial, we present the current landscape of quantum-resistant protocols with the focus on code-based and lattice-based systems. We start by introducing McEliece-like and Alekhnovich‘s public-key cryptosystems. Then, we cover the theory of lattices and its applications in learning with errors (LWE) and ring LWE. We present a security reduction for the Alekhnovich’s system and a worst-case to average-case reduction for LWE. Finally, the tutorial covers attacks in post-quantum cryptography and ends with a comparison of code-based and lattice-based cryptography with practical examples from current Crystals-Kyber (2022) and HQC (2025) standards adopted by the NIST.

Presenters: Shirin Saeedi Bidokhti, Hamed Hassani and Eric Lei

Lossy data compression is a well-established problem in information theory and signal processing with optimal yet computationally challenging solutions such as vector quantization, and practical yet sub-optimal solutions based on transform coding. Learning-based compression has emerged as a successful solution for compressing real-world data, especially for settings where reconstructions that are perceptually similar to the source are desired. Since then, a joint effort from the information theory and machine learning communities has been advancing the theory and practice of neural compression.

This tutorial provides an overview of the recent progress of the topic. In the first part of our tutorial, we study the fundamental limits of lossy compression and introduce neural estimation methods that can compute these limits for high dimensional sources using the power of generative models. These methods illustrate that recent neural compressors are sub-optimal. We next discuss channel simulation (also known as reverse channel coding), which sheds light on the structure of optimal solutions, but are of high complexity. In the third part of the tutorial, we build on these insights to discuss neural compressors that approach optimality yet remain low-complexity through the use of lattice coding techniques.

Presenters: Si-Hyeon Lee and Cheuk Ting Li

This tutorial provides a unified framework for addressing key problems in network information theory through three distinct yet interconnected parts. In Part 1, we introduce a general network setup that unifies canonical problems in network information theory, including source coding, channel coding, joint source-channel coding, and coding for computing. Then, we propose a unified asymptotic achievability bound, which is useful for deriving achievability bounds without detailed error analysis. In Part 2, we shift to the practical side, presenting the PSITIP software, a powerful tool for automating theorem proving in network information theory. By simplifying expressions involving entropy, mutual information, and Markov chains, PSITIP aids researchers and students in deriving capacity regions and bounds with ease. Part 3 explores a framework for deriving one-shot and finite-blocklength results using the Poisson matching lemma, which simplifies analyses typically relying on i.i.d. random codebooks. This part highlights the utility of Poisson processes as codebooks for nonasymptotic results, essential in short-packet communication scenarios.