Shohinur Pervez Shohan

Energy Measurement of Compressed Vision Models — Phase 1

Independent research, 2026 — paper in preparation for Sustainable Computing: Informatics and Systems

A paired RAPL (CPU package-0) + NVML (GPU) energy-measurement harness, built and run entirely on my own laptop (Intel i5-13450HX, RTX 3050).

Method

ResNet-18, MobileNetV3-Small, and EfficientNet-B0 on CIFAR-10, under INT8 post-training quantization and 30/50/70% structured pruning.

Main result

x86-CPU block, 6 sessions, 144 conditions, 1,440 windows, batch size 1: MAC-ratio reductions overstated realized CPU energy savings by 2.28x on average (95% CI 2.26–2.30).

Instrument validity

Found the NVML cumulative energy counter over-reports active GPU power by ~30% and switched to sampled power; also found RAPL package-0 and psys were double-counted in my own code, corrected offline from raw data.

Accuracy

Only ResNet-18 under INT8 kept ≥99% of reference accuracy. All 9 zero-finetune pruned models failed, some near chance.

Provenance

Analysis plan frozen in a git tag before data collection; Zenodo archive deposited after data collection. 40 deviations logged.

Energy Measurement of Compressed Deep Learning Models: A Literature Review

Submitted to Artificial Intelligence Theory and Applications (AITA, DergiPark), October 2026 — under review

An independent literature review examining the measurement layer beneath model compression's energy-savings claims, drawing on a purposively assembled corpus of 50+ works (2019–2026).

• Synthesises evidence that software energy estimators and hardware counters disagree by margins comparable to the energy savings compression claims to deliver.
• Shows FLOPs-based proxies correlate weakly with measured energy once memory bandwidth, batch size, and backend kernel realisation are accounted for.
• Finds that compression rankings established on one hardware platform do not reliably transfer to another.
• Sets out a research agenda around three gaps — measurement-tool validation, the nominal-versus-realised compression gap, and cross-platform rank-order preservation — that motivates the Phase 1 study above.

Hardware-Validated Energy-Aware Reliability Evaluation of Compressed Neural Networks

Proposed Master's research (2027 intake)

A deployment-aware framework bridging efficient and trustworthy machine learning: measures reliability (calibration, robustness) and hardware-validated energy from the same inference passes, then compares compression methods at matched measured energy rather than matched FLOPs.

• Tests whether software energy estimators (CodeCarbon, pyJoules) stay accurate on compressed models — existing validation studies stop at full-precision workloads, leaving quantisation's altered instruction mix and pruning's irregular memory access untested.
• Reconciles contradictory findings on whether compression improves or harms model calibration and robustness, via a controlled sweep across quantisation, structured pruning, unstructured pruning, and distillation with per-class disaggregation.
• Compares compression configurations within matched measured-energy bands instead of matched FLOPs or compression ratio, testing whether the ranking by accuracy agrees with the ranking by reliability (hypotheses fixed in advance, Kendall's tau).
• Builds on the completed Phase 1 research above (RTX 3050, RAPL/NVML); this proposal is scoped around resources not independently accessible — physical power meters, heterogeneous accelerators, and larger model families.
• Extends the same scepticism toward unvalidated measurement that the applicant's undergraduate thesis applied to spatial validation — and builds directly on the literature review above.

Multi-Decadal Land-Cover Mapping in the Chittagong Hill Tracts Using Spatially Blocked Classification, Area-Adjusted Accuracy Assessment, and Forest-Definition Sensitivity Analysis

Submitted to ICCIT 2026 (double-blind peer review) — currently under review

Originated as B.Sc. thesis (2024, graded A+), then independently extended in 2026 into a full conference paper without institutional funding or supervision.

Research question

Has forest coverage in the Chittagong Hill Tracts changed over three decades (1993–2023)?

Method

Machine-learning classification of six Landsat epochs, comparing Support Vector Machine (78.04%), Random Forest (77.10%), and CART (72.90%) under spatially-blocked holdout validation — a stricter design than a random split, chosen to avoid inflating accuracy through spatial autocorrelation. Benchmarked against the Hansen Global Forest Change dataset. Full pipeline built with free, open-source tools (Google Earth Engine, Python, QGIS).

Key finding

Combined forest area increased 4.4% (+20,693 ha) between 1993 and 2023 — but Dense Forest specifically declined by 8,235 ha, pointing to a shift in forest composition rather than simple loss or gain.

Validation

Area-adjusted 2023 classification accuracy of 73.17% (95% CI: 64.34–82.00%), plus a forest-definition sensitivity analysis to test how much the result depends on a single NDVI threshold. Full code and data pipeline released publicly for reproducibility.