AI Quality

ai-quality

Enabling AI Quality Control via Feature Hierarchical Edge Inference — Choi et al. (2022) (arXiv:2211.07860, 2022)

What this evaluates

Evaluates a feature-hierarchical edge inference framework's ability to dynamically allocate communication and computation resources to maximize AI quality under strict latency and energy constraints.

Datasets

• YOLO v3 (unspecified dataset) — total ?; splits: test (-1)

Metrics

• AI quality (mAP) (primary) — range: [0, 1]
  • Mean Average Precision (mAP) scaled by a feature size coefficient (δ_s = 12 mAP/Mbyte). The optimization objective maximizes the sum of AI quality across all mobiles.

Input / output format

Input: Raw input data (100 Kbytes), channel state information, and per-mobile computation/energy/latency constraints.

Output: Optimized uplink/downlink bandwidth allocation, edge/mobile computation FLOPS allocation, and selected feature scale, yielding a final AI quality score.

Scoring recipe

# Pseudo-code for the optimization objective

total_ai_quality = 0
for mobile in mobiles:
    feature_size = select_feature_scale(mobile.channel_state)
    ai_quality[mobile] = feature_size * delta_s  # delta_s = 12 mAP/Mbyte
    total_ai_quality += ai_quality[mobile]

# Constraints: avg_energy <= 5 J, avg_latency <= 15 sec
# Return allocation that maximizes total_ai_quality

Common pitfalls

• The paper uses simulation parameters rather than a standard benchmark dataset; YOLO v3 is mentioned but the underlying dataset is not specified.
• AI quality is a composite metric derived from mAP and feature size, not a direct classification accuracy score.
• Benchmarks compare against simplified resource allocation strategies (constant quality, computation-only optimization) rather than modern edge inference frameworks.

Evidence (verbatim from paper)

We consider two benchmarks: constant AI quality and FHEI with computation resource optimization only. For the first benchmark, every mobile’s AI quality is fixed but optimized under the same constraints as the proposed FHEI... δ_s | Coefficient of quality function | 12 mAP/Mbyte

Citation

@misc{choi2022enabling,
  title={Enabling AI Quality Control via Feature Hierarchical Edge Inference},
  author={Choi et al. (2022)},
  year={2022},
  note={arXiv:2211.07860}
}

• arXiv: 2211.07860