Protocol
Stochastic Optimization beyond SGD
Write an optimizer comparison that can reject a supposedly faster method once its true compute budget is counted.
45–80 active hours60–90 min sessionsAdvanced
Mastery contract
Specialization · Standard 1.0 · Write an optimizer comparison that can reject a supposedly faster method once its true compute budget is counted.
Loading your learning records…
Reasoning — not yet passed
Derive the variance-reduction estimator and explain when its expectation matches the full gradient.
Reliable implementation — not yet passed
Implementations match a hand-worked update and count every full/minibatch gradient evaluation.
Reproducible experiment — not yet passed
Compare SGD and variance reduction on the same objective and initialization; retain all seeds, objective gaps, gradient-evaluation counts, wall time and peak memory.
Defense and handoff — not yet passed
Use a measured convex-versus-neural comparison and an explicit failed assumption to delimit which variance-reduction findings transfer.
All four criteria must pass. Activity completion does not satisfy them.
An independent reviewer reruns your work and varies something you did not rehearse: a fresh input, a different working directory, or a declared edge case.
Pilot decisions are provisional, not external credentials. For an appeal or sensitive artifact, contact your designated pilot operator with the submission ID. Export your assessment history.
What you will do
6 total
Derive unbiasedness of a minibatch gradient and an SVRG-style control variate; state smoothness/convexity assumptions separately from neural experiments.
Implement SGD, momentum and a variance-reduced optimizer on a finite-sum convex objective with matching gradient counters.
Compare 3 optimizers at matched gradient-evaluation budgets over 5 seeds; repeat with an ill-conditioned feature transform.
Debug an adaptive optimizer with inconsistent bias correction and a benchmark that excludes full-gradient computation from cost.
Write a 500–900 word technical report linking the reasoning, code, measurements, and limitations; include the project decision below.
Assess all 4 mastery criteria against saved artifacts, request an independent review, and repeat each failed criterion on a fresh example.
Essentials
A reading session alone never satisfies a mastery criterion.
Store source references, assumptions, and artifact paths beside every result.
Keep an untouched check case that differs from the worked example.
Record all attempts, including failures and results that contradict your prediction.
Use the stated comparison conditions; document every deviation before drawing a conclusion.
Protocol guardrails
Do
- +State the expected result before running the comparison.
- +Keep one minimal reproducible failing case when debugging.
- +Record environment versions and the exact command used.
- +Compare explanations with saved intermediate values.
- +Ask a reviewer to challenge the weakest assumption.
Don't
- ×Do not copy a worked solution and present it as an independent implementation.
- ×Do not tune against held-out evaluation outcomes.
- ×Do not report only the best seed or discard inconvenient runs.
- ×Do not equate elapsed hours or a completed run with a passed assessment.
- ×Do not conceal reduced-scale experiments behind claims about the original full-scale result.
Protocol authorship
Written by NuthinButta as instructional design. The teaching sources are the official references linked in each milestone; the exercises, workload and pass thresholds are ours, not their authors'.