pn Thumb sm dsc 0642 2 scaled
Philipp Hennig (Project leader)
Max Planck Research Group Leader
pn Thumb sm thumb maren
Maren Mahsereci
Ph.D. Student
pn Thumb sm profile
Lukas Balles
Ph.D. Student
8 results

2017


Thumb md teaser image
Follow the Signs for Robust Stochastic Optimization

Balles, L., Hennig, P.

arXiv preprint arXiv:1705.07774, 2017 (article)

Abstract
Stochastic noise on gradients is now a common feature in machine learning. It complicates the design of optimization algorithms, and its effect can be unintuitive: We show that in some settings, particularly those of low signal-to-noise ratio, it can be helpful to discard all but the signs of stochastic gradient elements. In fact, we argue that three popular existing methods already approximate this very paradigm. We devise novel stochastic optimization algorithms that explicitly follow stochastic sign estimates while appropriately accounting for their uncertainty. These methods favorably compare to the state of the art on a number of benchmark problems.

link (url) Project Page [BibTex]


Thumb md early stopping teaser
Early Stopping Without a Validation Set

Mahsereci, M., Balles, L., Lassner, C., Hennig, P.

arXiv preprint arXiv:1703.09580, 2017 (article)

Abstract
Early stopping is a widely used technique to prevent poor generalization performance when training an over-expressive model by means of gradient-based optimization. To find a good point to halt the optimizer, a common practice is to split the dataset into a training and a smaller validation set to obtain an ongoing estimate of the generalization performance. In this paper we propose a novel early stopping criterion which is based on fast-to-compute, local statistics of the computed gradients and entirely removes the need for a held-out validation set. Our experiments show that this is a viable approach in the setting of least-squares and logistic regression as well as neural networks.

link (url) Project Page [BibTex]


Thumb md teaser
Coupling Adaptive Batch Sizes with Learning Rates

Balles, L., Romero, J., Hennig, P.

In Thirty-Third Conference on Uncertainty in Artificial Intelligence (UAI), (Editors: Gal Elidan and Kristian Kersting), 2017 (inproceedings) Accepted

Abstract
Mini-batch stochastic gradient descent and variants thereof have become standard for large-scale empirical risk minimization like the training of neural networks. These methods are usually used with a constant batch size chosen by simple empirical inspection. The batch size significantly influences the behavior of the stochastic optimization algorithm, though, since it determines the variance of the gradient estimates. This variance also changes over the optimization process; when using a constant batch size, stability and convergence is thus often enforced by means of a (manually tuned) decreasing learning rate schedule. We propose a practical method for dynamic batch size adaptation. It estimates the variance of the stochastic gradients and adapts the batch size to decrease the variance proportionally to the value of the objective function, removing the need for the aforementioned learning rate decrease. In contrast to recent related work, our algorithm couples the batch size to the learning rate, directly reflecting the known relationship between the two. On three image classification benchmarks, our batch size adaptation yields faster optimization convergence, while simultaneously simplifying learning rate tuning. A TensorFlow implementation is available.

Code link (url) Project Page [BibTex]

Code link (url) Project Page [BibTex]

2015


Thumb md maren ls
Probabilistic Line Searches for Stochastic Optimization

Mahsereci, M., Hennig, P.

In Advances in Neural Information Processing Systems 28, pages: 181-189, (Editors: C. Cortes, N.D. Lawrence, D.D. Lee, M. Sugiyama and R. Garnett), Curran Associates, Inc., 2015 (inproceedings)

Abstract
In deterministic optimization, line searches are a standard tool ensuring stability and efficiency. Where only stochastic gradients are available, no direct equivalent has so far been formulated, because uncertain gradients do not allow for a strict sequence of decisions collapsing the search space. We construct a probabilistic line search by combining the structure of existing deterministic methods with notions from Bayesian optimization. Our method retains a Gaussian process surrogate of the univariate optimization objective, and uses a probabilistic belief over the Wolfe conditions to monitor the descent. The algorithm has very low computational cost, and no user-controlled parameters. Experiments show that it effectively removes the need to define a learning rate for stochastic gradient descent. [You can find the matlab research code under `attachments' below. The zip-file contains a minimal working example. The docstring in probLineSearch.m contains additional information. A more polished implementation in C++ will be published here at a later point. For comments and questions about the code please write to mmahsereci@tue.mpg.de.]

Matlab research code link (url) Project Page [BibTex]

2015

Matlab research code link (url) Project Page [BibTex]

2013


Thumb md thumb hennigk2012 2
Quasi-Newton Methods: A New Direction

Hennig, P., Kiefel, M.

Journal of Machine Learning Research, 14(1):843-865, March 2013 (article)

Abstract
Four decades after their invention, quasi-Newton methods are still state of the art in unconstrained numerical optimization. Although not usually interpreted thus, these are learning algorithms that fit a local quadratic approximation to the objective function. We show that many, including the most popular, quasi-Newton methods can be interpreted as approximations of Bayesian linear regression under varying prior assumptions. This new notion elucidates some shortcomings of classical algorithms, and lights the way to a novel nonparametric quasi-Newton method, which is able to make more efficient use of available information at computational cost similar to its predecessors.

website+code pdf link (url) Project Page [BibTex]

2013

website+code pdf link (url) Project Page [BibTex]


Fast Probabilistic Optimization from Noisy Gradients

Hennig, P.

In Proceedings of The 30th International Conference on Machine Learning, JMLR W&CP 28(1), pages: 62–70, (Editors: S Dasgupta and D McAllester), 2013 (inproceedings)

PDF Project Page [BibTex]

PDF Project Page [BibTex]

2012


Thumb md thumb hennigk2012
Quasi-Newton Methods: A New Direction

Hennig, P., Kiefel, M.

In Proceedings of the 29th International Conference on Machine Learning, pages: 25-32, ICML ’12, (Editors: John Langford and Joelle Pineau), Omnipress, New York, NY, USA, July 2012 (inproceedings)

Abstract
Four decades after their invention, quasi- Newton methods are still state of the art in unconstrained numerical optimization. Although not usually interpreted thus, these are learning algorithms that fit a local quadratic approximation to the objective function. We show that many, including the most popular, quasi-Newton methods can be interpreted as approximations of Bayesian linear regression under varying prior assumptions. This new notion elucidates some shortcomings of classical algorithms, and lights the way to a novel nonparametric quasi-Newton method, which is able to make more efficient use of available information at computational cost similar to its predecessors.

website+code pdf link (url) Project Page [BibTex]

2012

website+code pdf link (url) Project Page [BibTex]