Pre-Register and Run a Small Neural-Network Ablation Study
Overview
What this challenge is about.
You will study how three architectural and regularization choices (depth: 2/4/8 hidden layers; activation: ReLU vs. GELU; weight decay: 0 / 1e-4 / 1e-3) affect a small MLP's test accuracy on CIFAR-10 patches (or a similar published benchmark). Pre-register the experiment matrix and the analysis plan (no peeking at test set). Run all 18 cells with three random seeds each. Produce a paper-style 6-page report with figures and confidence intervals, and deliver a 25-minute lab readout. Honest negative results are graded the same as positive ones.
The Brief
What you'll do, and what you'll demonstrate.
Run and write up a pre-registered ablation study on a small neural network with honest reporting of effects and confidence intervals.
Earning criteria — what you'll demonstrate
- Pre-register an ablation study and stick to the plan
- Run experiments across multiple seeds for honest confidence intervals
- Write a paper-style report with figures that communicate effect sizes
- Present negative results without overselling them
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.
Research Scientist
Pre-registered ablations with honest negative results are the credential research-scientist hiring loops grade above raw benchmark numbers.
This challenge sharpens
- experiment-design
- ablation-study
- statistical-analysis
ML Researcher
Running clean multi-seed experiments and writing them up paper-style is the ML-researcher's day-one job at any lab.
This challenge sharpens
- neural-networks
- ablation-study
- pytorch
Applied AI Scientist
Discipline around effect sizes and confidence intervals separates applied AI scientists from people who only run sweeps.
This challenge sharpens
- regularization
- experiment-design
- statistical-analysis