tasks/ml/manifests/ML01.yaml

# ============================================================================
# T01 — Dropout regularization strategies for shallow MLPs
# ----------------------------------------------------------------------------
# Unlike paper_replication's P01-P07, the "synthesis" here frames a research
# QUESTION rather than a known paper's method. The model must design the
# experiment (conditions, metrics, datasets) — we only commit to what a
# competent study of this topic would include and what the rubric expects.
# ============================================================================

id: ML01
title: "Dropout regularization strategies on shallow tabular MLPs"
arxiv_id: null
venue: "ARC-Bench 2026"
paper_asset: null

# The "synthesis" plays the role of the upstream briefing: research question,
# background, why the question matters, what "a reasonable experiment" looks
# like. It deliberately does NOT pre-specify a single method to reproduce.
synthesis: |
  Dropout-style regularization is a cornerstone of neural-net training for
  overfitting control, but its impact on tabular classification with shallow
  MLPs remains subtle. Standard element-wise dropout is the default; spatial
  (feature-wise) dropout treats correlated input features as a block; MC /
  variational dropout keeps dropout active at test time and averages over
  samples. For tabular data with small feature counts and mild non-linearity,
  these variants can produce equal test-set accuracy yet very different
  probability calibration — a known gotcha that makes accuracy-only
  evaluation misleading.

  A credible CPU-scale study of this topic (a) compares at least three
  dropout variants against a no-dropout control on multiple sklearn
  classification benchmarks, (b) reports both accuracy and calibration
  metrics (ECE, NLL, Brier), (c) averages over ≥5 seeds with statistical
  tests, and (d) checks for ablation integrity (that different dropout
  variants actually produce different output distributions, not just the
  same network). The research question is: *does the choice of dropout
  variant matter more for calibration than for accuracy on tabular data?*

hypotheses:
  - id: H1
    statement: "Variational / MC dropout produces lower Expected Calibration Error (ECE) than standard element-wise dropout on at least 2 of 3 tabular classification benchmarks, averaged over ≥5 seeds."
    measurable: true
  - id: H2
    statement: "Test-set accuracy differs by <2 absolute percentage points between dropout variants on datasets where all methods exceed 95% accuracy — i.e. accuracy is non-discriminative and calibration is needed to distinguish methods."
    measurable: true
  - id: H3
    statement: "The no-dropout control is strictly worse than at least one dropout variant in ECE on at least 2 of 3 datasets."
    measurable: true

experiment_design:
  research_question: "Does the choice of dropout variant (standard / spatial / variational) matter more for probability calibration than for test accuracy on tabular classification with shallow MLPs?"
  conditions:
    - name: "no_dropout"
      description: "Shallow MLP (2 hidden layers, 64 units, ReLU) trained with no dropout. Control condition."
    - name: "standard_dropout_p30"
      description: "Same MLP with standard element-wise dropout at p=0.3 between hidden layers."
    - name: "standard_dropout_p50"
      description: "Same MLP with standard element-wise dropout at p=0.5."
    - name: "spatial_dropout_p30"
      description: "Same MLP with feature-blockwise dropout at p=0.3 applied to the input layer."
    - name: "mc_dropout_p30_T20"
      description: "Same MLP with standard p=0.3 dropout left active at test time; prediction is the mean of T=20 stochastic forward passes."
  baselines:
    - "no_dropout is the no-regularization baseline"
  metrics:
    - name: "test_accuracy"
      direction: "maximize"
      description: "Fraction correctly classified on held-out 20% test split, mean over 5 seeds."
    - name: "ece"
      direction: "minimize"
      description: "Expected Calibration Error (15 bins) on the test split, mean over 5 seeds."
    - name: "nll"
      direction: "minimize"
      description: "Mean per-example negative log-likelihood on the test split."
    - name: "brier"
      direction: "minimize"
      description: "Mean Brier score on the test split."
  datasets:
    - name: "breast_cancer"
      source: "sklearn.datasets.load_breast_cancer"
    - name: "wine"
      source: "sklearn.datasets.load_wine"
    - name: "digits"
      source: "sklearn.datasets.load_digits"
  compute_requirements:
    gpu_required: false
    estimated_wall_clock_sec: 240

rubric_path: "experiments/arc_bench/config/ml/rubrics/ML01.json"