Implicit bias of SGD in L2-regularized linear DNNs: One-way jumps from high to low rank.
Zihan WangArthur JacotPublished in: ICLR (2024)
Keyphrases
- low rank
- matrix factorization
- missing data
- convex optimization
- linear combination
- rank minimization
- singular value decomposition
- matrix completion
- high dimensional data
- semi supervised
- stochastic gradient descent
- high order
- trace norm
- kernel matrix
- low rank matrix
- matrix decomposition
- robust principal component analysis
- singular values
- minimization problems
- low rank matrices
- affinity matrix
- small number
- recommender systems
- support vector
- training data