Sensor Compression Eval

sensor-compression-eval

FastML Science Benchmarks: Accelerating Real-Time Scientific Edge Machine Learning — Duarte et al. (2022) (arXiv:2207.07958, 2022)

What this evaluates

Evaluates unsupervised lossy compression of irregular sensor data on radiation-hardened edge ASICs. It probes the ability to reconstruct high-granularity calorimeter images under extreme bandwidth and latency constraints.

Datasets

• CMS HGCal Trigger Data — total ?; splits: test (-1); repo https://github.com/fastmachinelearning/fastml-science

Metrics

• Energy Mover's Distance (EMD) (primary) — range: other
  • A distance metric comparing reconstructed vs original radiation patterns that penalizes misreconstructed energies less if they are spatially close to the original deposit.

Input / output format

Input: 48 trigger cells represented as 8-bit fixed-point values (384 bits total), derived from aggregated hexagonal sensor array energies.

Output: Compressed bitstream of either 144 bits (moderate) or 48 bits (extreme), which is subsequently decompressed to reconstruct the 48 trigger cell values.

Scoring recipe

# EMD calculation between original (O) and reconstructed (R) 48-cell arrays
emd = compute_energy_movers_distance(O, R)
# Lower EMD indicates better reconstruction fidelity
# Report EMD alongside latency, area (mm2), and power (mW) constraints

Common pitfalls

• Must account for hexagonal sensor geometry during reconstruction, not standard rectangular grids.
• Compression budget is strictly 144b or 48b; exceeding this violates the benchmark constraint.
• Latency and power/area constraints are evaluated on the encoder ASIC, not just the algorithm.

Evidence (verbatim from paper)

The energy mover’s distance (EMD) is used to compare the reconstructed radiation patterns, giving smaller penalties for misreconstructed energies that are close-by to the original deposit.

Citation

@misc{duarte2022fastml,
  title={FastML Science Benchmarks: Accelerating Real-Time Scientific Edge Machine Learning},
  author={Duarte et al. (2022)},
  year={2022},
  note={arXiv:2207.07958}
}

• arXiv: 2207.07958