I am attaching the technical write-up for my selected Phase 1 submission, 326960.
The submitted estimator uses a private-safe Monte Carlo approach tailored to wide, deep, bias-free ReLU networks. Its design combines centered and whitened Gaussian samples, antithetic
input pairs, exact algebraic simplifications in the first layer, pilot-based identification of inactive neurons, conservative treatment of stable final-layer units, and selectively applied
fast matrix multiplication.
Several implementation choices were important for the final result. Antithetic symmetry avoids recomputing part of the first-layer work, the empirical Gram matrix is handled using its
symmetry, inactive activation rows are removed from later computations, and two-level Strassen/Winograd recursion is used only for sufficiently large products where it improves the total
FLOP-plus-runtime cost. Units whose behavior is uncertain remain on the fully sampled path.
The implementation uses only the official flopscope.numpy numerical interface. It does not rely on public-instance identification, lookup tables, MLP-specific routing, custom native
kernels, or unmetered computation. Sampling is sized using a conservative no-pruning cost estimate to preserve budget safety.
The attached PDF describes the estimator, the mathematical identities behind it, the computational tradeoffs, ablation results, failed experiments, limitations, and the progression leading
to the selected submission.
LLMs played a substantial role in brainstorming, coding, analysis, reviewing, and drafting. I led the research process, decided which experiments to pursue, checked the results, and made
the final decisions. I have not assumed that every idea or line of code was independently authored by me.
Attachment
phase1_writeup_submission_326960.pdf (136.2 KB)