WOLFE: A Declarative Machine Learning St...

Performing machine learning with existing toolkits on large datasets is quite a frustrating experience: each toolkit focuses on its own subclass of machine learning techniques, have their own different interfaces to how much of the underlying system is surfaced to the user, and don't support the iterative development that is required to tune machine learning algorithms and achieve satisfactory predictors. In this talk we present Wolfe, a declarative machine learning stack consisting of three crucial components: (1) Language: a math-like syntax embedded in Scala to concisely specify arbitrarily complex machine learning systems that unify most existing, and future, techniques, (2) Interpreter that transforms the declarative description into efficient code that scales to large-datasets, and (3) REPL: A new iPython-like IDE for Scala that supports the unique features for machine learning such as visualizing structured data, probability distributions, and state of optimization. (joint work with Sebastian Riedel, Tim Rocktaschel at UCL)