Speaker
Description
Machine learning methods are routinely used to explore large data sets, handle web searches, recommend products, perform voice recognition, recognize features in pictures and videos, and in countless other applications. Neural networks have been around for a long time. But in recent times the pervasive availability of huge amount of data, associated to large numbers of hidden parameters and to advancements in computing, storage and network technologies, has led to a rapid development of methods to train, optimize and exploit machine and deep learning techniques.
In scientific domains, resources such as data sets and computing or storage facilities are typically distributed across multiple providers, often located in different countries. The resources might be dedicated or opportunistic, and could be offered through worldwide computing Grids, via private or public Cloud infrastructures, via single data centers, or through a composition of the above.
This plethora of opportunities in accessing and exploiting the resources needed by deep learning methods brings about a number of issues. In this keynote, Davide Salomoni from the Italian National Institute for Nuclear Physics (INFN) will discuss how a "Deep learning as a Service" spanning heterogeneous, public or private infrastructures, could be designed, drawing on the experience gained by several international projects and initiatives. In particular, the keynote will cover topics such as orchestration of data and computing resources, data placement, automated instantiation of deep learning solutions and publication of reusable, trained data sets into catalogues.