Dictionary Learning

• Given : N unlabeled data points $x_i\in {\mathcal{R}}^d$
• To find:
  • Linear rep of these points based on set of basis vectors
    • $x={\Sigma }_i^k{r}_i\cdot d_i=Dr$
    • dimension d
    • r are repr weights corresponding to basis vector d
    • D is a dict with basis vectors
    • R contains weights. Scalar
    • $$||d_i|| \leq 1$
• Sparse Dictionary Learning Loss
• After learning, these can be used as discriminative Features
  • Expensive to compute