Task 3: Training

• Use off-the-shelf optimizer: cvxopt
• Calculate Q, p, G, h, A, b using training data.
• Understand those params.
• Unit test: Test if code is working or not.

Reference: Bishop's Chapter 7 (Sparse Kernel Machines ~ quadratic programming problem).

Task 5: Visualization + Comparison with sklearn

• Visualize hyperplane (linear, kernelized)
• Visualize support vectors.

Task 2: Kernelize SVM

• An interface for SVM with different kernel functions.
• Supported kernels follow Mercer's theorem: positive semi-definite matrix (e.g., Radial Basis Function)
• Create testing program for svm kernels

Task 4: Implement a custom optimizer for training

• SMO (Sequential Minimizing Optimizer): Bishop's + cs 229
• Recommend read: Numerical Optimization

Task 1: Inference

• Keep training data during inference (different from regression/classification: keep weight)
• Define public API for SVM: store param; given param how to inference to get the prediction
• Later on, keep only the support vectors.