Presented By O’Reilly and Intel AI
Put AI to work
8-9 Oct 2018: Training
9-11 Oct 2018: Tutorials & Conference
London, UK

Harden and improve your deep learning models with targeted ensembles

Alan Mosca (nPlan | Birkbeck, University of London)
11:55–12:35 Wednesday, 10 October 2018
Models and Methods
Location: Windsor Suite
Secondary topics:  Deep Learning models, Ethics, Privacy, and Security

Who is this presentation for?

  • Machine Learning engineers and researchers

Prerequisite knowledge

  • A basic understanding of deep learning (e.g., convolutional networks, LSTMs, and backpropagation)

What you'll learn

  • Explore tools to improve the robustness of models, both against adversarial attacks and to increase accuracy

Description

Adversarial attacks on deep neural networks are starting to show up in a lot of literature. There are even examples of how to defeat a neural network by changing just the value of one pixel. Ensemble methods and distillation have been shown to be great tools for deploying models into production. Unfortunately, until now, some of these methods did not work well for deep learning. New methods have emerged that enable users to create and train deep learning ensembles that can be distilled into a fast single model, which can be hardened against adversarial attacks.

Alan Mosca leads you on a journey from a single model through all the various techniques that enable the construction of production-ready deep learning models. You’ll explore ensemble methods, distillation, adversarial training, and some novel optimization techniques, which have all been shown to help improve accuracy and increase robustness. Alan then shows you how these techniques can be used in the Toupee deep learning framework to create production-ready models.

Photo of Alan Mosca

Alan Mosca

nPlan | Birkbeck, University of London

Alan Mosca is the cofounder and CTO of nPlan and a part-time doctoral researcher at Birkbeck, University of London, where his research focuses on deep learning ensembles and improvements to optimization algorithms in deep learning. Previously, Alan worked at Wadhwani Asset Management, Jane Street Capital, and several software companies as well as on several consulting projects in machine learning and deep learning.

Leave a Comment or Question

Help us make this conference the best it can be for you. Have questions you'd like this speaker to address? Suggestions for issues that deserve extra attention? Feedback that you'd like to share with the speaker and other attendees?

Join the conversation here (requires login)