ISIT 2022
204 papers
The 2022 IEEE International Symposium on Information Theory (ISIT) featured 238 papers addressing core information-theoretic problems and their applications in machine learning. A prominent research theme involved generalization in deep learning, with studies focusing on computable generalization bounds for self-certified learning, neural collapse under cross-entropy loss, and the behavior of overparameterized models. Optimization techniques also received attention, specifically Bregman proximal methods and strategies for reducing gradient variance in stochastic processes. Another area of inquiry was partial information decomposition for analyzing complex, multimodal, and multivariate systems.
The year's contributions also covered advancements in coding theory and signal processing, including list decoding of binary alternant codes, universal 1-bit compressive sensing, and universal error correction using foundation models. Research further explored practical communication constraints, such as finite block-length regimes and joint communication and privacy requirements. Prolific authors for this edition included Yihan Zhang, Bo Bai, Antonia Wachter-Zeh, and Deniz Gündüz. The proceedings reflected a continued integration of information theory with federated learning and the mitigation of spurious correlations in large-scale models.
Top Research Directions
Research directions most represented by ISIT 2022 papers.
Top Authors
Impactful Papers
Papers from ISIT 2022 that shaped research clusters.
Foundational
Privacy Leakage in Discrete-Time Updating Systems
foundational
Balancing Information Freshness and Priv
Influential
Partial Information Decomposition via Deficiency for Multivariate Gaussians
influential
Scaling Multivariate Information Decompo
Tighter Expected Generalization Error Bounds via Convexity of Information Measures
influential
Tightening the Bounds Between Total Vari
MIMO Systems with One-bit ADCs: Capacity Gains using Nonlinear Analog Operations
influential
Beyond the High SNR Asymptotic Assumptio
Optimizing Estimated Directed Information over Discrete Alphabets
influential
Probabilistic Constellation Shaping
DeepNP: Deep Learning-Based Noise Prediction for Ultra-Reliable Low-Latency Communications
influential
Overcoming the Noiseless Feedback Assump
Nonparametric Matrix Estimation with One-Sided Covariates
influential
Generalizing Matrix and Tensor Recovery
DeepNP: Deep Learning-Based Noise Prediction for Ultra-Reliable Low-Latency Communications
influential
Modeling Temporal Dependencies with Recu