07-29, 19:00–19:30 (UTC), Red Track
Machine Learning is more than large complex models, it is moving towards integrating existing domain knowledge to better inform learning processes. See how Flux expresses that problem in the modern machine learning paradigm.
Flux.jl has been evolving with a host of improvements from the ground up. A major change from last year is that we have officially launched a stable release that uses Zygote.jl as its AD package, opening up a lot more of the ecosystem to take advantage of it. We will show how well it plays with Julia’s existing state-of-the-art packages enabling more kinds of modelling than ever, and show how the community is taking advantage of it.
We will also showcase Flux’s new APIs and features that open up the awesome open-source community to express more complex ideas and bring domain skills into their ML stories. We will show how we’ve worked on improving performance through a myriad changes across the ecosystem, while keeping it easy to use as ever.
We would also discuss the design patterns that enable users to write elegant and performant Julian code that can best take advantage of the powerful automatic differentiation capabilities that we have built over time.
Dhairya is currently a Data Scientist at Julia Computing Inc. and maintains and develops Julia's ML stack with Flux and Zygote.