Label noise (stochastic) gradient descent implicitly solves the Lasso for quadratic parametrisation.
Loucas Pillaud-VivienJulien ReygnerNicolas FlammarionPublished in: CoRR (2022)
Keyphrases
- stochastic gradient descent
- label noise
- least squares
- regularization parameter
- active learning
- loss function
- matrix factorization
- step size
- training data
- random forests
- image restoration
- pairwise
- cross validation
- support vector machine
- objective function
- online algorithms
- feature selection
- weight vector
- model selection
- convergence rate
- sample size
- cost function
- multiple kernel learning
- linear svm
- supervised learning
- multi class
- decision trees