
Already know some of this? A 2-minute test skips the lessons you know.
Take the placement testA free 12-lesson course built from MLSquare's 1h 47m of YouTube video. 1 day at 90 minutes a day, with study notes and a quiz in every lesson. A 2-minute placement test skips what you already know.
Taught by MLSquare on YouTube. Videos play in YouTube's own player, so the creator keeps every view.
Review the linear algebra prerequisites for PCA including covariance matrix construction, eigenvalue/eigenvector properties, trace and determinant relationships, and eigendecomposition computation.
Understand the PCA objective of minimizing reconstruction error, its equivalence to maximizing variance, and the step-by-step algorithm: center data, compute covariance matrix, find top K eigenvectors as principal components.
Learn why Kernel PCA captures nonlinear relations by mapping to a higher-dimensional space, and how the kernel matrix K = ΦᵀΦ relates to the kernel function k(xᵢ, xⱼ) = φ(xᵢ)ᵀφ(xⱼ).
Understand kernel matrix centering (Kc = K - 1ₙK - K1ₙ + 1ₙK1ₙ), the criteria for valid kernels (symmetric, positive semidefinite), and popular kernels: polynomial (xᵢᵀxⱼ + c)ᵖ and RBF exp(-||xᵢ-xⱼ||²/2σ²).
Compute principal components in Kernel PCA via eigendecomposition of the centered kernel matrix Kc, and project new data points using kernel evaluations without explicitly computing φ.
Apply MLE by writing the likelihood function, taking the log, differentiating to find critical points, and verifying maxima with the second derivative test.
Perform Bayesian estimation by combining prior and likelihood to obtain the posterior, leveraging conjugate priors (e.g., Beta-Binomial) for closed-form posterior computation.
Define GMMs as mixtures of K Gaussians with parameters πₖ, μₖ, Σₖ, and compute the number of free parameters (3K-1 for 1D with shared variance).
Derive the EM algorithm for GMM: E-step computes responsibilities γ(zᵢₖ) = πₖN(xᵢ|μₖ,Σₖ)/ΣⱼπⱼN(xᵢ|μⱼ,Σⱼ); M-step updates πₖ, μₖ, Σₖ using weighted averages.
Solve a concrete GMM/EM problem: compute P(component=1 | x₁=1) with given π, μ, σ², simplifying using cancellations to obtain 1/(1+e⁻²).
Understand K-means as iterative minimization of within-cluster sum-of-squares: initialize assignments, recompute centroids as means, reassign points to nearest centroid, repeat until convergence to a local optimum.
Apply K-means++ initialization: pick first centroid uniformly at random, then choose subsequent centroids with probability proportional to squared distance from nearest existing centroid.