Presentation Title

wwu_tinker : A free hyperparameter tuning tool

Presentation Type

Poster

Abstract

While machine learning model parameters can be learned from a set of training data, training machine learning models almost always requires setting "hyperparameters" that cannot be learned from the data (e.g. the order of a polynomial or the number of hidden layers in a deep neural network). Accurately tuning these hyperparameters can be the difference between success and failure for the learned model. Naive hyperparameter tuning strategies are problematic: hand tuning requires too much human intervention and can miss promising regions, while brute force "grid search" ends up sinking too much time exploring unpromising regions of the hyperparameter space. These flaws are particularly pronounced for large machine learning problems, where the number of hyperparameters can exceed 10 and it can take hours, days or weeks to train a single model. While there currently exist a handful of industry-oriented services that implement similar optimization algorithms, these tend to be expensive, underdeveloped, or prohibitively complex. wwu_tinker is a web-service solution with an extensible API that researchers can use to tune hyperparameter configurations both efficiently and flexibly. The service we provide will be free, well-documented, easy to use, and will be geared towards smaller research labs that do not need all the bells and whistles provided by enterprise services.

Start Date

10-5-2018 12:00 PM

End Date

10-5-2018 2:00 PM

This document is currently not available here.

Share

COinS
 
May 10th, 12:00 PM May 10th, 2:00 PM

wwu_tinker : A free hyperparameter tuning tool

While machine learning model parameters can be learned from a set of training data, training machine learning models almost always requires setting "hyperparameters" that cannot be learned from the data (e.g. the order of a polynomial or the number of hidden layers in a deep neural network). Accurately tuning these hyperparameters can be the difference between success and failure for the learned model. Naive hyperparameter tuning strategies are problematic: hand tuning requires too much human intervention and can miss promising regions, while brute force "grid search" ends up sinking too much time exploring unpromising regions of the hyperparameter space. These flaws are particularly pronounced for large machine learning problems, where the number of hyperparameters can exceed 10 and it can take hours, days or weeks to train a single model. While there currently exist a handful of industry-oriented services that implement similar optimization algorithms, these tend to be expensive, underdeveloped, or prohibitively complex. wwu_tinker is a web-service solution with an extensible API that researchers can use to tune hyperparameter configurations both efficiently and flexibly. The service we provide will be free, well-documented, easy to use, and will be geared towards smaller research labs that do not need all the bells and whistles provided by enterprise services.