Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Model-Aware Rate-Distortion Limits for Task-Oriented Source Coding

Andriy Enttsel, Vincent Corlay

TL;DR

This work reframes task-oriented source coding (TOSC) through indirect rate–distortion (iRD) theory, showing that identifiability-based oracle bounds can be misleading in realistic settings. It derives task-model-aware bounds by incorporating suboptimal task models and transmitter constraints, introducing three approaches: E&C, iE&C, and sample–and–communicate (S&C). Through theoretical bounds and classifications benchmarks (MNIST, CIFAR-100, ImageNet), it demonstrates that current learned TOSC methods fall short of these limits, primarily due to transmitter complexity rather than coding inefficiency. The findings emphasize the need to jointly optimize rate–distortion and computational/architectural budgets to achieve practically near-optimal TOSC performance.

Abstract

Task-Oriented Source Coding (TOSC) has emerged as a paradigm for efficient visual data communication in machine-centric inference systems, where bitrate, latency, and task performance must be jointly optimized under resource constraints. While recent works have proposed rate-distortion bounds for coding for machines, these results often rely on strong assumptions on task identifiability and neglect the impact of deployed task models. In this work, we revisit the fundamental limits of single-TOSC through the lens of indirect rate-distortion theory. We highlight the conditions under which existing rate-distortion bounds are achievable and show their limitations in realistic settings. We then introduce task model-aware rate-distortion bounds that account for task model suboptimality and architectural constraints. Experiments on standard classification benchmarks confirm that current learned TOSC schemes operate far from these limits, highlighting transmitter-side complexity as a key bottleneck.

Model-Aware Rate-Distortion Limits for Task-Oriented Source Coding

TL;DR

This work reframes task-oriented source coding (TOSC) through indirect rate–distortion (iRD) theory, showing that identifiability-based oracle bounds can be misleading in realistic settings. It derives task-model-aware bounds by incorporating suboptimal task models and transmitter constraints, introducing three approaches: E&C, iE&C, and sample–and–communicate (S&C). Through theoretical bounds and classifications benchmarks (MNIST, CIFAR-100, ImageNet), it demonstrates that current learned TOSC methods fall short of these limits, primarily due to transmitter complexity rather than coding inefficiency. The findings emphasize the need to jointly optimize rate–distortion and computational/architectural budgets to achieve practically near-optimal TOSC performance.

Abstract

Task-Oriented Source Coding (TOSC) has emerged as a paradigm for efficient visual data communication in machine-centric inference systems, where bitrate, latency, and task performance must be jointly optimized under resource constraints. While recent works have proposed rate-distortion bounds for coding for machines, these results often rely on strong assumptions on task identifiability and neglect the impact of deployed task models. In this work, we revisit the fundamental limits of single-TOSC through the lens of indirect rate-distortion theory. We highlight the conditions under which existing rate-distortion bounds are achievable and show their limitations in realistic settings. We then introduce task model-aware rate-distortion bounds that account for task model suboptimality and architectural constraints. Experiments on standard classification benchmarks confirm that current learned TOSC schemes operate far from these limits, highlighting transmitter-side complexity as a key bottleneck.
Paper Structure (18 sections, 34 equations, 4 figures)

This paper contains 18 sections, 34 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Rate--distortion limits of single-task TOSC
  3. Operational bounds via factorized coding strategies
  4. Compress-and-estimate
  5. Estimate-and-compress
  6. Identifiability and oracle rate--distortion function
  7. Motivating example
  8. Task model-dependent bounds
  9. Classification
  10. E&C
  11. iE&C
  12. Sample & communicate
  13. Generalizability
  14. Numerical results
  15. Rate-distortion
  16. ...and 3 more sections

Figures (4)

  • Figure 1: Gaussian mixture model with overlapping classes: probability density (left) and rate–distortion performance (right).
  • Figure 2: Qualitative behavior of the rate--distortion bounds.
  • Figure 3: Rate--distortion performance across standard classification benchmarks based on common vision task models.
  • Figure 4: State-of-the-art accuracy versus bits-per-pixel performance, compared with theoretical bounds and a naive E&C baseline.