Hasty Briefsbeta

Bilingual

The linear algebra and calculus behind every model

6 hours ago
  • Every machine learning model relies on three pillars: linear algebra, calculus, and convex optimization.
  • The MNIST dataset is treated as a matrix of 70,000 rows (images) and 784 columns (pixels) after flattening.
  • Data exploration shows balanced classes, no missing values, and a bimodal pixel intensity distribution.
  • Vectors have L1 and L2 norms; dot products measure similarity between images.
  • Matrix multiplication forms a Gram matrix of pairwise similarities.
  • Broadcasting allows adding a scalar bias to tensors without explicit loops.
  • Projections decompose images into components along and perpendicular to the mean image direction.
  • The Singular Value Decomposition (SVD) factors the centered data matrix and reveals 626 significant singular values out of 784.