Cnn Edge Optimization Eval

cnn-edge-optimization-eval

A Comparative Study of CNN Optimization Methods for Edge AI: Exploring the Role of Early Exits — Fernandez et al. (2026) (arXiv:2604.14789, 2026)

What this evaluates

Evaluates the trade-offs between predictive accuracy, model compression, and dynamic inference efficiency of CNN optimization techniques (pruning, quantization, early-exit) for edge deployment. It probes how different architectures handle static compression versus input-adaptive latency reduction under hardware-constrained conditions.

Datasets

• Unspecified classification dataset — total ?; splits: test (-1)

Metrics

• accuracy (%) (primary) — range: percent
  • Percentage of correctly classified samples relative to the total number of evaluation samples.
• compression ratio — range: other
  • Ratio of the baseline model size to the optimized model size, indicating static memory footprint reduction.
• early-exit rate (%) — range: percent
  • Proportion of samples routed through an early-exit branch instead of the final classification layer, indicating dynamic computation skipping.
• label loyalty (%) — range: percent
  • Percentage of samples where the optimized model's prediction matches the baseline model's prediction, measuring decision consistency.
• avg. inference time (ms) — range: other
  • Mean time per inference measured on target edge hardware (CPU/CUDA), capturing real-world latency.

Input / output format

Input: Image inputs processed by CNN architectures (ResNet-152, EfficientNet-B2, MobileNet-V2, ShuffleNet-V2) with applied optimization techniques (structured pruning, PTQ/DQ quantization, or early-exit branches).

Output: Class predictions, confidence scores for early-exit routing decisions, and hardware performance logs (latency, speed-up, CPU/GPU utilization, RAM usage).

Scoring recipe

def compute_metrics(predictions, gold, baseline_preds, baseline_size_mb, opt_size_mb, inference_times):
    accuracy = sum(p == g for p, g in zip(predictions, gold)) / len(predictions)
    compression = baseline_size_mb / opt_size_mb
    ee_rate = sum(1 for p in predictions if p.exit_branch != 'final') / len(predictions)
    label_loyalty = sum(p == b for p, b in zip(predictions, baseline_preds)) / len(predictions)
    avg_time = sum(inference_times) / len(inference_times)
    return accuracy, compression, ee_rate, label_loyalty, avg_time

Common pitfalls

• Assuming pruning yields consistent compression across all architectures; it degrades significantly on efficiency-optimized nets like EfficientNet-B2 and MobileNet-V2.
• Confusing accuracy-optimal (acc-opt) and inference-optimal (inf-opt) early-exit configurations, which have opposite effects on accuracy and early-exit rate.
• Evaluating early-exit benefits solely on static metrics like model size, ignoring the critical runtime latency and memory savings measured on edge hardware.

Evidence (verbatim from paper)

Table 2 summarizes the results across all architectures and configurations, reporting model size (MB), compression ratio (relative to the baseline), accuracy (%), label loyalty (%), and probability loyalty (%). For models incorporating early exits, the table additionally reports the early-exit rate (%), which represents the proportion of samples that exit through an early-exit branch rather than the final-exit layer.

Citation

@misc{fernandez2026cnnedge,
  title={A Comparative Study of CNN Optimization Methods for Edge AI: Exploring the Role of Early Exits},
  author={Fernandez et al. (2026)},
  year={2026},
  note={arXiv:2604.14789}
}

• arXiv: 2604.14789