Skip to content

Foundations Track

This track builds core ML execution discipline: math intuition, data understanding, feature preparation, robust splitting, and evaluation quality.

  1. projects/deep-learning-math-foundations-showcase
  2. projects/neural-network-foundations-showcase
  3. projects/pytorch-training-regularization-showcase
  4. projects/sota-supervised-learning-showcase
  5. projects/eda-leakage-profiling-showcase
  6. projects/feature-engineering-dimred-showcase
  7. projects/sota-unsupervised-semisup-showcase

Core Skills Covered

  • Vectors, matrices, gradients, entropy, and optimization traces.
  • Perceptrons, activations, backpropagation, and initialization choices.
  • PyTorch tensors, training loops, optimizers, schedulers, and regularization tradeoffs.
  • Train/validation/test discipline.
  • Stratified, group-aware, and time-aware split strategies.
  • Univariate and bivariate EDA.
  • Missingness diagnostics and leakage checks.
  • Feature encoding, selection, and dimensionality reduction.

Evidence Artifacts To Inspect

  • artifacts/vector_operations.csv
  • artifacts/gradient_descent_trace.csv
  • artifacts/activation_comparison.csv
  • artifacts/decision_boundary_summary.csv
  • artifacts/optimizer_comparison.csv
  • artifacts/regularization_ablation.csv
  • artifacts/splits/split_manifest.json
  • artifacts/eda/univariate_summary.csv
  • artifacts/eda/bivariate_vs_target.csv
  • artifacts/leakage/leakage_report.csv
  • artifacts/selection/selection_scores.csv

Suggested Reflection Prompts

  • Which split strategy is most defensible for this dataset and why?
  • Where does model capacity start to help, and where does it start to overfit?
  • Which optimizer or regularizer improved learning without hiding what the model was doing?
  • Which feature engineering step improved generalization most?
  • Which leakage check would fail first if pipeline order changed?