Eyal Fishel .Subject : AI-Aided Inference under Communication Constraints

Eyal Fishel Ph.D. Student, Electrical Engineering supervision: Nir Shlezinger :Abstract A broad range of technologies relies on remote inference, wherein data acquired is conveyed over a communication channel for inference on a remote server. Communication often occurs over rate-limited channels, requiring compression to reduce latency. While deep learning enables joint design of compression, encoding, and inference rules, existing mechanisms are typically static and struggle to adapt to dynamic links or varying channel conditions. To address this, we propose a learned, rate-adaptive, and privacy-aware vector quantization mechanism tailored for remote inference over varying communication links. Our approach, Adaptive Rate Task-Oriented Vector Quantization (ARTOVeQ), leverages nested codebooks trained via progressive learning and an augmented VQ-VAE architecture. By integrating mutual information optimization, the method simultaneously maximizes task-relevant information and suppresses private data while enabling successive refinement for low-latency inference. This allows multiple resolutions to be used concurrently when conveying high-dimensional data, ensuring consistency across bit levels. Numerical results demonstrate that the proposed scheme supports multiple rates, adheres to the data processing inequality, achieves rapid inference that improves with additional bits, and approaches the performance of single-rate deep quantization methods. These results highlight the potential of combining learned quantization with task- and privacy-aware information-theoretic objectives to achieve efficient, adaptive, and privacy-preserving remote inference.