Two Talks: Factored Bandits for online learning and Aha moments in deep learning at Zendesk
The schedule will look as follows:
17:00 Meet and greet!
17:15 Julian Zimmert: Factored Bandits for online learning
18:15 Chris Hausler: Aha moments in deep learning at Zendesk
19:00 Hang out
19:30 Doors closing
Aha moments in deep learning at Zendesk
Chris Hausler and Arwen Griffioen discuss Zendesk’s experience with deep learning, using the example of Answer Bot, a question-answering system that resolves support tickets without agent intervention. They cover the benefits Zendesk has already seen and challenges encountered along the way.
Answer Bot uses deep learning to understand customer queries, responding with relevant knowledge base articles that allow customers to self-serve. Research and development behind the ML models underpinning Answer Bot has been rewarding but punctuated with pivotal deviations from the charted course: deep learning was not Zendesk’s first approach. Chris and Arwen walk you through the journey from product ideation to traditional ML approaches with per-customer models to the current release that utilizes word embeddings and recurrent neural networks to provide a single global model that can serve tens of thousands of accounts.
Defining the problem space and the metrics Zendesk wanted to optimize for
How the team approached the problem with traditional ML
Why they chose to take a big bet and pivot to deep learning
Getting the team up to speed with deep learning and TensorFlow
Why deep learning is great
Why deep learning is great—but not magic
Processes and frameworks for experimentation, iteration, and validation
Factored Bandits for Online Learning
The Factored Bandits model is an online learning framework for learning with limited (bandit) feedback, where actions can be decomposed into a Cartesian product of atomic actions. Bandit algorithms have been successfully applied in problems such as online marketing, ranker evaluation or adaptive routing in networks. However, many of the algorithms used in practice make strong assumptions on the underlying reward distribution. When these assumptions are violated, as they typically are in practice, theoretical performance guarantees are lost. With Factored Bandits, we aim to close the gap between theory and practice by weakening the assumptions of the model. We introduce the Temporary Elimination Algorithm that achieves (up to constants) optimal performance under the weakened assumptions of Factored Bandits.
Hosted by Henrik Brink and Christian Igel