Galen Andrew
ABSTRACT
The L-BFGS limited-memory quasi-Newton method is the algorithm of choice for optimizing the parameters of large-scale log-linear models with L2 regularization, but it cannot be used for an L1-regularized loss due to its non-differentiability whenever some parameter is zero. Efficient algorithms have been proposed for this task, but they are impractical when the number of parameters is very large. We present an algorithm Orthant-Wise Limited-memory Quasi-Newton (OWL-QN), based on L-BFGS, that can efficiently optimize the L1-regularized log-likelihood of log-linear models with millions of parameters. In our experiments on a parse reranking task, our algorithm was several orders of magnitude faster than an alternative algorithm, and substantially faster than L-BFGS on the analogous L2-regularized problem. We also present a proof that OWL-QN is guaranteed to converge to a globally optimal parameter vector.
- Benson, J. S., & More, J. J. (2001). A limited memory variable metric method for bound constraint minimization.Google Scholar
- Bertsekas, D. P. (1999). Nonlinear programming. Athena Scientific.Google Scholar
- Byrd, R. H., Lu, P., Nocedal, J., & Zhu, C. Y. (1995). A limited memory algorithm for bound constrained optimization. SIAM Journal on Scientific Computing, 16, 1190--1208. Google ScholarDigital Library
- Charniak, E., & Johnson, M. (2005). Coarse-to-fine n-best parsing and maxent discriminative reranking. ACL. Google ScholarDigital Library
- Collins, M. (2000). Discriminative reranking for natural language parsing. ICML (pp. 175--182). Google ScholarDigital Library
- Darroch, J., & Ratcliff, D. (1972). Generalised iterative scaling for log-linear models. Annals of Mathematical Statistics.Google ScholarCross Ref
- Efron, B., Hastie, T., Johnstone, I., & Tibshirani, R. (2004). Least angle regression. Annals of Statistics.Google Scholar
- Gao, J., Andrew, G., Johnson, M., & Toutanova, K. (2007). A comparative study of parameter estimation methods for statistical NLP. ACL.Google Scholar
- Goodman, J. (2004). Exponential priors for maximum entropy models. ACL.Google Scholar
- Kazama, J., & Tsujii, J. (2003). Evaluation and extension of maximum entropy models with inequality constraints. EMNLP. Google ScholarDigital Library
- Lee, S.-I., Lee, H., Abbeel, P., & Ng, A. (2006). Efficient L1 regularized logistic regression. AAAI-06.Google Scholar
- Malouf, R. (2002). A comparison of algorithms for maximum entropy parameter estimation. CONLL. Google ScholarDigital Library
- Minka, T. P. (2003). A comparison of numerical optimizers for logistic regression (Technical Report). Microsoft Research.Google Scholar
- Ng, A. Y. (2004). Feature selection, L1 vs. L2 regularization, and rotational invariance. ICML. Google ScholarDigital Library
- Nocedal, J., & Wright, S. J. (1999). Numerical optimization. Springer.Google Scholar
- Perkins, S., & Theiler, J. (2003). Online feature selection using grafting. ICML.Google Scholar
- Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B.Google Scholar
- Zhu, C., Byrd, R. H., Lu, P., & Nocedal, J. (1997). Algorithm 778: L-bfgs-b: Fortran subroutines for large-scale bound-constrained optimization. ACM Trans. Math. Softw., 23, 550--560. Google ScholarDigital Library
- Scalable training of L1-regularized log-linear models
Recommendations
An inexact successive quadratic approximation method for L-1 regularized optimization
We study a Newton-like method for the minimization of an objective function $$\phi $$ that is the sum of a smooth function and an $$\ell _1$$ℓ1 regularization term. This method, which is sometimes referred to in the literature as a proximal Newton ...
An Augmented Lagrangian Method for $\ell_{1}$-Regularized Optimization Problems with Orthogonality Constraints
A class of $\ell_1$-regularized optimization problems with orthogonality constraints has been used to model various applications arising from physics and information sciences, e.g., compressed modes for variational problems. Such optimization problems are ...
Nonmonotone Barzilai---Borwein Gradient Algorithm for $$\ell _{1}$$ℓ1-Regularized Nonsmooth Minimization in Compressive Sensing
This study aims to minimize the sum of a smooth function and a nonsmooth $$\ell _{1}$$ ℓ 1 -regularized term. This problem as a special case includes the $$\ell _{1}$$ ℓ 1 -regularized convex minimization problem in signal processing, compressive sensing, machine learning, data mining, ...
Comments