Ziv Aharoni

Research

My research spans information theory, communication, and machine learning. My long-term goal is to translate the principles that make communication systems both computationally efficient and close to fundamental statistical limits into the design of AI algorithms. I aim to use these principles to reduce the computational complexity and energy footprint of learning and generative models while retaining strong statistical performance.

Conservation-law curves tracking exact ground truth across three diffusion noise paths
Local information derivatives integrate to global cross-entropy.

01 · Generative models

Information-theoretic foundations of diffusion

Information-theoretic characterization of diffusion models for memoryless noise processes.

We show that data–model cross-entropy can be written as an integral of local information-theoretic derivatives for a broad class of memoryless noise processes. The result unifies likelihood characterizations for discrete and continuous diffusion and identifies the noise channel as a consequential design choice.

  • Global likelihood from local posterior quantities.
  • A common treatment of discrete and continuous diffusion models.

Featured papers

  • Conservation Laws for Diffusion Models 2026

    Establishes conservation laws that express data–model cross-entropy as an integral of local information-theoretic derivatives, providing a unified likelihood characterization for discrete and continuous diffusion models.

Paper Project Code
Neural polar decoding architecture and performance comparison
Structured neural decoding preserves the recursive polar-code architecture.

02 · Reliable communication

Learning-based coding for unknown channels

Data-driven code design and structured neural decoding for unknown channels with and without memory.

I develop neural polar decoders that learn from black-box channel observations while retaining the recursive structure of polar coding. The same framework supports decoding, mutual-information estimation, and code-rate or input-distribution optimization.

  • Unknown channels with and without memory.
  • Wireless receivers for 5G channels.
  • Deletion channels and DNA data storage.

Featured papers

  • Neural Polar Decoders for Receivers in Wireless Communication 2026

    Adapts neural polar decoding to practical OFDM and single-carrier systems. Across 5G channels, it improves error rates and throughput while supporting rate matching, higher-order modulation, and operation without pilots.

  • Neural Polar Decoders for DNA Data Storage 2025

    Develops low-complexity neural polar decoders for DNA storage channels with insertions, deletions, and substitutions, with evaluations on synthetic channels, multiple reads, and real Nanopore data. On Nanopore data, the decoder matches the performance of a state-of-the-art transformer using only 3% of its parameters.

Optimized polar codes DNA storage Software
Directed-information neural estimation and optimization workflow
Neural estimators turn directed information into a trainable objective.

03 · Capacity estimation

Neural information estimation and sequential optimization

Data-driven capacity estimation and optimization for channels with memory and feedback.

I use neural estimators and reinforcement learning to compute directed information, optimize input processes, and expose capacity-achieving structure in channels with memory. The resulting methods apply across discrete and continuous spaces.

  • Directed-information estimation from samples.
  • Optimization over sequential input processes.
  • Feedback-capacity computation and duality bounds.

Featured papers

Directed information Feedback capacity All publications