webzuweb/philosophy-as-inductive-bias

Philosophy as Inductive Bias

A dependency-light PyTorch reference implementation accompanying the paper:

Philosophy as Inductive Bias: Five Under-Formalized Epistemic Mechanisms and Differentiable Objectives for Deep Learning Viacheslav Golitsyn (webzuweb), Independent Researcher, Moscow, 2026. Full paper included: paper.pdf · source paper.tex, refs.bib.

What this repository is

It turns five ideas from epistemology and philosophy of cognition into explicit, differentiable objectives, each with unit tests and documented failure modes:

1. Epoche — temporarily bracket a learned prior and measure what remains evidence-driven.
2. Hermeneutic circle — iteratively reconcile part and whole representations toward a fixed point.
3. Dialectical synthesis (Aufhebung) — synthesize conflicting representations while provably preserving information from both parents.
4. Epistemic virtues — jointly train calibration, humility, honesty w.r.t. available support, openness to counter-evidence, and abstention.
5. Peircean semiosis — recursively model a sign–object–interpretant triad with reconstruction and anti-collapse regularization.

A differentiable multi-hypothesis abductive objective is included as a baseline.

What this repository is not

• Not a claim of state-of-the-art performance.
• Unit tests validate software behavior and gradient flow, not scientific effectiveness.
• The philosophical labels do not establish that human cognition uses these exact mechanisms.
• Each objective should be compared against strong, compute-matched baselines and may honestly lose.

Files

philosophia/               # the package
  __init__.py              # public API (see below)
  epoche.py                # EpocheRegularizer, epoche_penalty
  hermeneutic.py           # HermeneuticConsistency, hermeneutic_loss
  dialectic.py             # DialecticalSynthesis, aufhebung_step
  perspectivism.py         # PerspectivalEnsemble, perspectival_disagreement
  virtue.py                # VirtueRegularizer, expected_calibration_error
  abduction.py             # AbductiveScorer, abductive_selection_loss
  wrapper.py               # PhilosophiaWrapper
tests/test_philosophia.py  # 23 unit + gradient-flow tests
examples/demo_calibration.py
paper.pdf                  # 10-page paper, 30 references
paper.tex, refs.bib        # arXiv-ready source
requirements.txt
LICENSE                    # MIT

Install & run

pip install -r requirements.txt          # just torch
python -m pytest tests/ -q               # 23 passed
python examples/demo_calibration.py

Quick use

import torch
from philosophia import PhilosophiaWrapper, EpocheRegularizer, VirtueRegularizer

# each objective is a small nn.Module returning a differentiable scalar penalty
epoche = EpocheRegularizer(prior_uniformity_weight=0.25)
virtue = VirtueRegularizer()

Public API: EpocheRegularizer, epoche_penalty, HermeneuticConsistency, hermeneutic_loss, DialecticalSynthesis, aufhebung_step, PerspectivalEnsemble, perspectival_disagreement, VirtueRegularizer, expected_calibration_error, AbductiveScorer, abductive_selection_loss, PhilosophiaWrapper.

Citation

@misc{golitsyn2026philosophy,
  title  = {Philosophy as Inductive Bias: Five Under-Formalized Epistemic
            Mechanisms and Differentiable Objectives for Deep Learning},
  author = {Golitsyn, Viacheslav},
  year   = {2026},
  note   = {Reference implementation with unit tests}
}

License

Code: MIT (see LICENSE). The paper text remains the author's; a publication venue may apply its own license separately.