Benchmarks#
These notebooks reproduce experiments from
Runje & Shankaranarayana (2023) using
mononet. They are committed with their outputs and re-executed manually
before each release — see
CONTRIBUTING.md.
Each notebook also benchmarks against airtai/monotonic-nn
(the paper’s original PyTorch reference) installed at notebook-execution
time via --no-deps (see tools/execute-benchmarks.sh).
Sections#
Protocol — how we train, select, and report; and why our numbers differ from the original papers.
Overview — high-level summary.
Reproducing the paper — per-dataset notebooks and summary tables for all five benchmark datasets from the ICML 2023 paper.
Flavor comparison — Phase 2a Optuna HP-search results comparing the four
mode × residualflavors (mixed/split×plain/residual).Deep residual accuracy — does the now-trainable depth (residual
sub_depth=2skips) improve real-dataset test accuracy over the shallow tuned flavors?Loan size-ladder — does deep monotone residual win once the dataset is large enough?
Large-dataset screen — max-size deep vs shallow on the benchmark roster; gates each dataset to a full ladder study or the standard benchmark by the Δ criterion.
Deep-network init — the static
mixedinit that fixes moderate-depth trainability, and the deep-depth limitation it does not solve.Alternate base result — tuned shallow (≤4 layers)
alternateflavor vs. the best ofsplit/mixed, on the five paper datasets.HP-search sensitivity — how each flavor’s tuned result evolves with Optuna trial count; search-saturation and meta-overfitting diagnostics reconstructed from the committed study storage.