Quantize a CNN for Battery-Powered Wildlife Cameras at a Climate Nonprofit
Overview
What this challenge is about.
You receive an FP32 CNN (MobileNetV2 fine-tuned to 22 species, around 13 MB) and a hold-out test set of 4,000 images. Quantize to int8 (post-training quantization first, then quantization-aware training if accuracy regresses). Measure accuracy retention, model size, latency, and energy per inference on an emulated Cortex-M target (use the TensorFlow Lite Micro tooling or a documented x86 proxy with a documented scaling factor). Recommend a final configuration.
The Brief
What you'll do, and what you'll demonstrate.
Quantize the species classifier to fit the 4 MB / 200 ms / 30 mJ budget while keeping top-1 accuracy within 2 percentage points of the FP32 baseline.
Earning criteria — what you'll demonstrate
- Apply post-training and quantization-aware training methods on real CNNs
- Reason about size/latency/energy budgets jointly
- Run on-device-style profiling (or documented proxy)
- Communicate quantization trade-offs to a firmware audience
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
MLOps Engineer
Edge-deployment quantization with explicit energy budgets is the kind of work MLOps engineers do at hardware-product or conservation-tech companies.
This challenge sharpens
- quantization
- edge-deployment
- energy-budgeting
Machine Learning Engineer
QAT and on-device inference profiling are increasingly part of the MLE job description on edge-AI teams.
This challenge sharpens
- quantization
- qat
- edge-deployment
Computer Vision Engineer
Shipping vision models inside tight hardware budgets is a recurring CV-engineer responsibility on hardware-product teams.
This challenge sharpens
- edge-deployment
- evaluation
- quantization