[X-POST] Learning Algorithms for Machine Learning on Big Data

RSVP here https://www.meetup.com/DataScience/events/234595976/ and here https://www.meetup.com/MachineLearning/events/234595775/

Title:

There Are Much Better Algorithms For Machine Learning On Huge Datasets

Abstract:

In nearly all fields of science and engineering, the amount of data available is growing at unprecedented rates. Applications no longer produce data sizes from megabytes to gigabytes, but rather from terabytes to petabytes (and beyond). Machine learning is one the key tools we use to make sense of these ever-growing quantities of data. We now use machine learning methods every day; they are behind software for e-mail spam filtering, product and advertisement recommendation systems, Microsoft's Kinect, Google Translate, speech recognition on phones, and now self-driving cars. The successes and potential of machine learning are driving the need to develop techniques that can consider even larger datasets and more complicated models.

A major challenge is that the "learning" in most machine learning models involves solving a numerical optimization problem, and standard numerical optimization codes are simply not up for the task of fitting very-complicated models to huge data sets. The default way to address this challenge is to use "stochastic gradient" methods. Instead of repeatedly going through your entire dataset between each model update, these methods alternate between looking at small random parts of the data and updating the model. These methods have been enormously successful, but they are enormously frustrating to use: it can be very hard to tune their parameters, to decide when to stop running them, and even if you address these issues they still converge very slowly. In this talk I'll give an overview of these methods, and then discuss a revolution that is happening in numerical optimization and machine learning with the development of a new class of stochastic gradient methods in 2012. Not only do the new methods make tuning parameters and deciding when to stop much easier, but these algorithms are dramatically faster than the old algorithms both in theory and practice.

About Mark Schmidt:

Mark Schmidt (http://www.cs.ubc.ca/%7Eschmidtm/) is a professor at UBC whose focus is on Machine learning, Numerical Optimizaiton, Probabilistic Graphical Models, and Causality.

Schedule:

• 6:30PM Doors are open, feel free to mingle
• 7:00 Presentation start
• ~8:15 Off to a nearby restaurant for food, drinks, and breakout discussions

Getting There:

By transit there a number of high frequency buses (check Google Maps or the Translink site for your particular case) that will get you there. For the drivers, there is a fair bit of street parking (free and pay) in the area, especially after 6.