Bayesian Optimization | Doing Stats and ML Wrong

This meetup is sponsored by Intel. We will have free drinks and food.

Talk 1: Vu Pham

Title: Bayesian Optimization for Hyper-parameter tuning and beyond

Abstract: Most Machine Learning practitioners spend relentless efforts on feature engineering and hyper-parameter tuning for their predictive models, yet the process is manual, time-consuming, and boring. This talk presents an overview of the problems and different approaches to tackle them. In particular, we will take a close look into Bayesian Optimization both in its theoretical background as well as usage in popular toolboxes. The results of using BayesOpt for training predictive models on some real-world datasets will be shown, and code will be provided. At a higher level of abstraction, the talk also gives a glimpse on an integrated framework for ML practitioners which can automate the boring parts of their job, help them to be more productive and creative.

Description: While random search works reasonably well for hyper-parameter tuning, recent advances in Bayesian Optimization made it practical and shown great results. We will review the theory and take a look into some popular implementations like Spearmint and SigOpt. Results of using those optimizers to tune predictive models on real-world datasets will be given. For feature engineering, we will show a generic pipeline inspired by recent research in the field, which allow features to be automatically generated and selected, then fed into training algorithms and the whole pipeline is tuned with BayesOpt. This generic approach, whenever done right, can be quite competitive in term of performance.

Bio: Vu is a machine learning engineer at SAP Innovation Center where he helps bring Machine Learning solutions to their customers. Prior to SAP, he was in some startups and research labs, trained Machine Learning models and built frameworks on top of Spark and CUDA, back in the days when there was no tensorflow. He spoke at Spark Summit and Strata+Hadoop World on several aspects of Machine Learning and Data Analytics systems.

Talk 2: Gerrit Gruben

Title: Doing statistics and ML wrong in the big data age

Abstract: Faking statistics or doing bogus research on data has always been a classic and interesting topic. In the big data age, we observe otherwise rare phenomena such as the Simpson's paradox more often. There are also limits to our methods, both theoretical - think "black swan" - and human - think biases. I want to touch several topics to increase your consciousness and sharpen your critical thinking as an ethical data scientist. As everyone in Machine Learning has created a faulty experimental design at least once, this presentation is also of a high practical value. I will show-case you concrete examples of where the model evaluation has been screwed up for the disadvantage of human beings.

Bio: Gerrit is a freelance data scientist with experience as a lead developer and data scientist for several companies based in Berlin. Together with data science retreat he trains data science talent for the market. Gerrit holds degrees in mathematics and computer science from KIT and FU Berlin and has worked in the area of machine learning for four years.https://www.linkedin.com/in/ggruben/ https://github.com/uberwach

Sponsor Note:

To receive invitations for Intel Machine Learning and Deep Learning webinars, news and tools register in the link below:
http://bit.ly/Berlinmeetup

This is a great platform to learn: How to access free resources and tools to turn you into an AI expert! What Intel tools, technology and frameworks to use to build for the performance and scale you need. Why developing on Intel Architecture can be a simplified AI strategy that’s built to last.

Get the latest information on tools, optimized frameworks, and training for artificial intelligence, machine learning, and deep learning through the new Metal Minds - Intel® Nervana™ AI Academy Newsletter (http://bit.ly/Berlinmeetup).