Applied research · NIT Puducherry · Dec 2023 — Jan 2024 · LIDC-IDRI · pruning · quantization
Energy-efficient lung nodule analysis
Lung nodule classification and segmentation models on the LIDC-IDRI CT dataset, redesigned to use a fraction of the energy and memory without losing accuracy.
What I did
Designed models that were about 77% more energy-efficient, with an 86% cut in compute use and no loss in accuracy.
Pruning, quantization and architecture changes cut memory and compute needs by up to 95%.
Who I made it with
NIT Puducherry, Department of CSE.
What it’s for
Done during my internship at NIT Puducherry’s CSE department, using pruning, quantization and architecture changes to make the models practical for real-time analysis.
Why it was hardIn medical imaging the usual move is a bigger model. The interesting constraint was the opposite: the same answer, with far less compute.
In numbers
- 77%
- more energy-efficient, with 86% less compute and no loss in accuracy. SourceAs reported on my previous portfolio. The baseline model and the hardware it was measured on aren’t recorded there.
Receipts
- Previous portfolio, as it stood before this redesign (self-reported)Self-reported: IIIT Dharwad (Data Science & AI), research at IIT Bombay Educational Technology, CRF approach for SSMR.
- LIDC-IDRI dataset (The Cancer Imaging Archive)Public lung CT dataset used for the NIT Puducherry work.
Next, a little further back in timeAffectBots →