Dates

Description

Courses

• [RKL] Representation and Kernel Learning - Fabio Aiolli
  Machine learning deals with techniques and algorithms that starting from empirical data allow a computer system to acquire new knowledge, or to correct and/or to refine knowledge already available. The success of machine learning algorithms heavily depends on data representation. Even the most performing classification algorithm will fail when provided with a bad input representation of data. In this course, the problem of data representation for machine learning will be introduced. In particular, we will focus on kernel machines (e.g. support vector machines for classification and regression), the theory behind these kinds of algorithms, and well-founded principled methods to learn the implicit representation of kernels from data. Topics include: basics of supervised learning theory, kernel based algorithms and SVM, kernel representation and kernel learning.
• [LC] Languages for Concurrency - Paolo Baldan
  Concurrency is pervasive in modern computing systems due to the success of web and network applications, and, at a lower level, of multicore and multiprocessor systems. The course is intended to provide a gentle introduction to the main concepts underlying the theory of concurrency and its impact on the design of languages for concurrent and distributed programming. We start by briefly reviewing some formal modelling languages for concurrency (Calculus of Communicating Systems) and the associated reasoning and automated verification techniques. Then we discuss the import of this theory on the design of languages with modern concurrency primitives, more abstract and manageable than the low-level model based on threads and locks, used by programmers for years for dealing with concurrency and parallelism. A tentative program follows:
  • Formal Specification languages for concurrency and tools for automated verification
  • Google Go and channel based concurrency
  • Erlang and the actor model for distributed concurrency 

  • Clojure, data concurrency for free using functional languages
• [RTS] Perspectives on Real-time Scheduling in the Wuthering Landscape of Parallel Computing - Tullio Vardanega
  Multicore processors are supplanting traditional processor architectures in virtually all application domains, promising considerably higher compute power within low-cost, into easily composable (scalable) commodity cluster units. This change much affects the real-time embedded systems domain, whose basic theory builds on the single-CPU assumption. This earlier challenge of real-time scheduling – the problem of assigning units of CPU time to units of computation (mains) so that all applications complete in time – then becomes a parallel problem, where multiple assignments occur simultaneously. Meanwhile, the embedding of compute-intensive services in critical infrastructures requires high-performance computing to seek timeliness. This need presents a new real-time scheduling problem, where application mains fork, multiple times, in a large number of very small units of computation that proceed in parallel until the next join point. In this mini-course we will look at these two facets of parallel real-time scheduling:
  • Dispatching multiple application mains to multiple CPUs, with no more than one CPU assigned to a single application at any one time.
  • Dispatching multiple short jobs from the same application to multiple CPUs multiple times, offering guarantees of timely completion, under contention from other similar applications.

Schedule

Organizers

Contacts