Overview
What this challenge is about.
You receive a trained 280 KB CNN keyword spotter (10 keywords + silence + unknown) with 96.1% top-1 accuracy on the Google Speech Commands test set. Apply structured pruning (channel-level) to two intermediate widths, then knowledge-distill a smaller student CNN (target under 50 KB INT8) using the pruned models as teachers. Report top-1 accuracy, false-accept rate at a 1% miss budget, and quantized SRAM footprint. Write a one-page memo on the effort-vs-accuracy curve to inform the next firmware cycle.
The Brief
What you'll do, and what you'll demonstrate.
Compress a keyword-spotting model into a 50 KB INT8 footprint via pruning + distillation while keeping accuracy production-viable.
Earning criteria — what you'll demonstrate
- Apply structured pruning to a real CNN keyword spotter
- Use knowledge distillation to compress a larger model into a smaller student
- Evaluate KWS models with the metrics that matter for product (FAR at fixed FRR)
- Communicate compression trade-offs to firmware leadership
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.
Machine Learning Engineer
Hitting a strict on-device SRAM budget through pruning + distillation is the daily job of an MLE on any consumer-audio or hearable team.
This challenge sharpens
- pruning
- knowledge-distillation
- model-compression
MLOps Engineer
Reproducible compression + deployment pipelines for embedded targets transfer directly to MLOps roles at edge-AI companies.
This challenge sharpens
- edge-inference
- quantization
- model-compression
Applied AI Scientist
Mapping a compression budget into an effort/accuracy memo for firmware leadership is the applied-AI-scientist craft of turning research into product decisions.
This challenge sharpens
- knowledge-distillation
- pruning
- pytorch