A PHP Error was encountered

Severity: 8192

Message: Non-static method URL_tube::usage() should not be called statically, assuming $this from incompatible context

Filename: url_tube/pi.url_tube.php

Line Number: 13

KDD 2017 | Applied Data Science Invited Talks - {title}

Applied Data Science Invited Talks

Szilard Pafka


Wednesday 1:30pm – 3:30pm, Room 200D

Title: Machine Learning Software in Practice: Quo Vadis?

Due to the hype in our industry in the last couple of years, there is a growing mismatch between software tools machine learning practitioners wish for, what they would truly need for their work, what’s available (either commercially or open source) and what tool developers and researchers focus on. In this talk we will give a couple of examples of this mismatch. Several surveys and anecdotal evidence show that most practitioners work most of the time (at least in the modeling phase) with datasets that t in the RAM of a single server, therefore distributed computing tools are very of- ten overkill. Our benchmarks (available on github [1]) of the most widely used open source tools for binary classification (various implementations of algorithms such as linear methods, random forests, gradient boosted trees and neural networks) on such data show over 10x speed and over 10x RAM usage difference between various tools, with ”big data” tools being the most inefficient. Significant performance gains have been obtained by those tools that incorporate various low-level (close to CPU and memory architecture) optimizations. Due to the hype in our industry in the last couple of years, there is a growing mismatch between software tools machine learning practitioners wish for, what they would truly need for their work, what’s available (either commercially or open source) and what tool developers and researchers focus on. In this talk we will give a couple of examples of this mismatch. Several surveys and anecdotal evidence show that most practitioners work most of the time (at least in the modeling phase) with datasets that in the RAM of a single server, therefore distributed computing tools are very of- ten overkill. Our benchmarks (available on github [1]) of the most widely used open source tools for binary classification (various implementations of algorithms such as linear methods, random forests, gradient boosted trees and neural networks) on such data show over 10x speed and over 10x RAM usage difference between various tools, with ”big data” tools being the most inefficient. Significant performance gains have been obtained by those tools that incorporate various low-level (close to CPU and memory architecture) optimizations. Nevertheless, we will show that even the best tools show degrading performance on the multi-socket servers featuring a high number of cores, systems that have become widely accessible more recently. Finally, while most of this talk is about performance, we will also argue that machine learning tools that feature high-level easy- to-use APIs provide increasing productivity for practitioners and therefore are preferable.

Speaker Bio

Szilard studied Physics in the 90s and has obtained a PhD by using statistical methods to analyze the risk of financial portfolios. Next he has worked in a bank quantifying and managing market risk. About a decade ago he moved to California to become the Chief Scientist of a credit card processing company doing everything data (analysis, modeling, data visualization, machine learning, data infrastructure etc). He is also the founder/organizer of several meetups in Los Angeles (R, data science etc) and the data science community website datascience.la. He is teaching data science and machine learning in graduate programs at CEU (Europe) and UCLA (California).