webzuweb/epoche-shortcut-experiment

Epoche: teaching one model to bracket a shortcut prior

A small, fully reproducible experiment on shortcut learning: a two-path classifier learns handwritten digits while a color prior is made correct in 99% of training examples. We then test whether the same model can be trained to temporarily "bracket" (switch off) that prior and fall back to evidence-only inference — an idea borrowed, loosely, from Husserl's epoche (ἐποχή, "suspension").

This repository contains the code, per-seed results, aggregated CSVs, bootstrap CIs and all figures behind the article. Everything runs offline on CPU in a few minutes.

• Data: sklearn.datasets.load_digits (identical to the HF mirror sklearn-docs/digits, 1,797 × 8×8 digits).
• Compute: CPU, PyTorch. 6 training schemes × 3 test environments × 10 seeds.
• Author: Viacheslav Golitsyn (webzuweb).

The setup in one paragraph

Each digit gets a 10-valued color prior. In training the color equals the label with probability ρ = 0.99. The model has an explicit evidence path (pixels) and an explicit prior path (color). We evaluate three passes: full (both inputs), bracketed (prior zeroed), prior-only (evidence zeroed), across three shifts: ID (ρ=0.99), Unbiased (ρ=0.10), Anti (ρ=0.00, the color always lies).

Headline results (10 seeds, 100 epochs)

Method / inference mode	ID acc, %	Unbiased acc, %	Anti acc, %
ERM, full input	95.75 ± 1.19	86.39 ± 1.61	86.00 ± 1.68
ERM, post-hoc bracket	88.28 ± 1.54	88.08 ± 1.34	88.86 ± 1.21
Evidence-only	89.08 ± 1.70	88.69 ± 1.45	89.31 ± 1.38
Prior randomization	89.06 ± 1.93	88.83 ± 1.55	89.42 ± 1.33
Epoche 2-view	90.83 ± 1.28	88.94 ± 1.29	89.31 ± 1.32
Epoche 3-view dual, normal	94.28 ± 1.36	88.06 ± 1.53	87.47 ± 1.48
Epoche 3-view dual, bracketed	88.61 ± 1.66	88.86 ± 1.57	89.50 ± 1.49
Epoche 3-view robust, normal	91.81 ± 1.25	88.94 ± 1.22	89.03 ± 1.25

Honest negative control. Simply zeroing the color branch of a plain ERM model after training already reaches 88.86% on Anti. The extra Epoche training adds only +0.64 pp over that post-hoc baseline (paired bootstrap 95% CI [+0.17; +1.17]).

The substantive result is not raw OOD accuracy but a switchable model: the dual configuration keeps +5.19 pp ID accuracy over evidence-only (95% CI [+4.44; +6.00]) while matching its robustness once the prior is explicitly bracketed.

Reproduce

pip install -r requirements.txt
python epoche_experiment.py --output-dir results --seeds 10 --epochs 100
# faster smoke test:
python epoche_experiment.py --output-dir results_smoke --seeds 2 --epochs 30

Extra analysis figures (forest plot, cliff, per-seed spread, heatmap):

python extra_charts.py

What's in this repo

epoche_experiment.py        # full experiment (model, losses, 6 methods, metrics, figures)
extra_charts.py             # 4 additional charts from the saved results
requirements.txt            # torch, numpy, scikit-learn, pandas, matplotlib
README_verification.md      # independent reproduction notes (numbers matched to 2nd decimal)
results/
  results.md                # summary table
  deployment_summary.csv    # per method/mode means + stds
  summary_results.csv       # all metrics, mean/std over seeds
  per_seed_results.csv      # raw per-seed metrics
  paired_bootstrap.json     # paired differences + 95% CIs
  config.json               # exact hyperparameters
  training_curve_seed0_*.csv
  figures/*.png             # 9 figures

Key hyperparameters

AdamW, lr=0.01, weight_decay=3e-4, 100 epochs, 60/20/20 stratified split, sensor noise σ=0.20. Dual weights (α,β,γ)=(0.10,0.02,0.02); robust (0.50,0.10,0.10). Full config in results/config.json.

Loss (three-view Epoche)

$$\mathcal{L}=\mathrm{CE}(z^{full},y)+\alpha \mathrm{CE}(z^{brk},y)+\beta JS(p^{full}\mathrm{\parallel }{p}^{brk})+\gamma D_{KL}(p^{prior}\mathrm{\parallel }{U}_{10})\mathrm{.}\backslash mathcal\; L\; =\; \backslash operatorname\{CE\}(z^\{full\},y)\; +\; \backslash alpha\backslash ,\backslash operatorname\{CE\}(z^\{brk\},y)\; +\; \backslash beta\backslash ,\; JS(p^\{full\}\backslash Vert\; p^\{brk\})\; +\; \backslash gamma\backslash ,\; D\_\{KL\}(p^\{prior\}\backslash Vert\; U\_\{10\}).$$

Intended use & limitations

This is a controlled diagnostic, not a benchmark win. It assumes the suspicious prior is known in advance (a dedicated branch). A single 8×8 toy benchmark cannot show that the method beats GroupDRO/JTT/ReBias on natural images — the planned next step is grodino/waterbirds with worst-group metrics. Use it to study the mechanism, teach the idea, or as a starting point for a stronger evaluation.

Citation

@misc{golitsyn2026epoche,
  title  = {Philosophy as Inductive Bias: Bracketing a Shortcut Prior (Epoche)},
  author = {Golitsyn, Viacheslav},
  year   = {2026},
  note   = {Reproducible experiment, sklearn-docs/digits}
}

References

• Geirhos et al. Shortcut Learning in Deep Neural Networks. Nature MI, 2020. https://doi.org/10.1038/s42256-020-00257-z
• Arjovsky et al. Invariant Risk Minimization. 2019. https://arxiv.org/abs/1907.02893
• Sagawa et al. Distributionally Robust Neural Networks for Group Shifts. ICLR 2020. https://arxiv.org/abs/1911.08731
• Bahng et al. Learning De-biased Representations with Biased Representations. ICML 2020. https://proceedings.mlr.press/v119/bahng20a.html
• Liu et al. Just Train Twice. ICML 2021. https://proceedings.mlr.press/v139/liu21f.html

License

MIT (see license in the metadata above).