DATAX397, 94.63%
RBFN benchmark and Bayesian inference
Python, NumPy, custom scikit-learn-style estimators, Optuna, Bayesian inference
Self-directed study with two tracks: a custom Radial Basis Function Network with a rigorous benchmark, and Bayesian inference theory.
- Built a custom AdaptiveRBFN (adaptive sigma, ridge regularisation, and a correlation-based centre-pruning heuristic using graph matching) that placed 2nd of 9 models by average rank across 16 Optuna-tuned datasets - ahead of tuned Random Forest, XGBoost, and MLP, and within 0.004 mean accuracy of a tuned SVM, at roughly 2.8x faster tuning than the naive baseline.
- Independently derived the connection between regularisation and Bayesian priors (Ridge as the MAP estimate under a Gaussian prior, Lasso under a Laplace prior), and wrote a from-scratch explanation of INLA (Integrated Nested Laplace Approximation).