Invited talks, 2011

General-purpose GPU computing uses graphics processors for work beyond graphics. Wikipedia’s GPGPU introduction explains that role in relation to computation traditionally assigned to CPUs. Convolutional neural networks add the subject of learned filters, connecting a model’s operations with the parallel resources that execute them.

Reconfigurable hardware allows the arrangement of computational logic to change after manufacturing. An FPGA expresses that idea through programmable logic blocks and interconnections. This provides background for the title’s pairing of dataflow and vision, where the organisation of operations matters alongside the representation of an image.

Pervasive data brings questions about the variety and arrival pattern of observations. Large collections can contain different types of information, while a continuous stream raises different organisational questions from a fixed batch. These are general data-management issues around the behavioural subject named in the title.

A learning tool provides an interface between a mathematical task and its execution. Its representation of inputs, models and operations shapes how a calculation can be expressed. This places software design within the workshop’s larger discussion of algorithms, systems and practical learning tasks.

Scientific learning problems can involve large collections of observations and complicated models. The 2011 workshop named astronomy among its application domains, and its interests extended to other computational sciences. The title placed fundamental-particle research within that broader discussion of data and learning.

A statistical application needs both a learning calculation and a way to organise its input. Relational data management supplies one familiar vocabulary through rows, columns and queries. The title connected these data-oriented concerns with the construction and maintenance of statistical applications.

A sketch stores a compact summary rather than every observation in its original form. Frequency sketches, for example, estimate counts from a stream using a data structure built around hash functions. This gives background for the title’s combination of sketches and real-time data.

The 2012 USENIX paper on resilient distributed datasets described a distributed-memory abstraction for fault-tolerant computation. Its motivating subjects included iterative algorithms and interactive data exploration, and it identified Spark as an implementation. This supplies a separate research reference for the cluster and in-memory vocabulary in the invited title.

MapReduce divides a calculation into operations over input and operations over grouped results. A distributed implementation also coordinates tasks, data transfers and fault tolerance. Hadoop is one setting for this programming model, connecting the title with the workshop’s systems vocabulary.

Josh Wills gave the talk listed for Jeff Hammerbacher.

The 2010 GraphLab paper described an abstraction for asynchronous iterative algorithms with sparse dependencies. It also addressed the consistency of data used by those computations. This supplies context for graph-oriented learning, where the relationships among values affect which work can proceed together.

Vowpal Wabbit is an online and interactive machine-learning system. Online learning concerns updates as observations are processed, rather than treating every task as a single fixed batch calculation. The tutorial title connects this computational pattern with a practical tool, while the optimisation page explains its general vocabulary.

Graph-oriented computation makes relationships among data part of the programming model. Local changes can depend on neighbouring values, so consistency and the scheduling of updates become useful concepts. The graph-parallel page explains those ideas alongside the GraphLab research line and other graph computation models.