The Robustness of SVM Kernels to Noise: A Comparative Analysis on MNIST

About this project

A study testing SVM kernel robustness (Linear, Polynomial, RBF) against noisy MNIST data[cite: 264]. While RBF leads on clean data at 96.50%, it degrades fastest under stress[cite: 266]. The Polynomial kernel proved most robust, maintaining nearly 40% accuracy even at 100% noise[cite: 267]. This project explores the bias-variance tradeoff and why highly flexible models can struggle with messy, real-world inputs[cite: 268, 272].