Data mining is the process of extracting information and patterns often previously unknown from large quantities of data using various techniques from areas such as machine learning, pattern recognition, and statistics. Data could be in files, relational databases, or other types of databases such as multimedia databases. Data may be structured or unstructured. This seminar is divided into three parts. Part I describes technologies for data mining. Without the underlying technologies such as database systems, warehousing, and machine learning, one cannot develop data mining as a technology area. We describe various supporting technologies and discuss their relationship data mining. Part II describes techniques and tools for data mining. It starts with a discussion of the various steps to data mining such as getting the data ready, carrying out the mining, pruning the results, and evaluating the outcomes and then provides an overview of the various data mining techniques. Part III describes emerging trends in data mining. These include mining distributed, heterogeneous, and legacy data sources, mining multimedia data such as text, images, and video, mining data on the world wide web, metadata aspects of mining, and the privacy issues of data mining.
Significant advances in the mathematical analysis of practical symmetric encryption schemes have been enabled by the modeling of block ciphers with secure finite families of pseudo-random functions (PRFs). However, such analyses sometimes ignore operational security concerns that date back to WWII, such as the presence of known and predictable text in encrypted messages, as well as more recent concerns of detecting message forgeries. In this presentation, I will describe an analysis method that addresses these operational concerns and retains the advantages of modeling block ciphers with PRFs. I will also address the question of whether the length of secret keys -- the only parameter considered important by political debates -- must be explicitly taken into account by analyses of encryption schemes.
Messaging is of growing importance in the business world. It provides a sound basis for the federated operation of many disparate applications working in heterogeneous environments. This is essential in a business world where the requirements of applications change much faster than the applications can be reworked. Typical business scenarios include mergers, customer based focus replacing departmentally organized data processing, and the addition of Web front ends to established applications.
Messaging can be seen as a version of traditional e-mail; application to application rather than person to person, and with messages often worth millions of dollars. Scalability, robust transactional behaviour and quick recovery from machine failure are essential.
General messaging has grown up over this decade; though there are specialized messaging applications dating back more than thirty years. Most work so far has been in message transport systems, ensuring delivery of a message from one application to another. This is now being complemented by higher level message brokers. Message brokers understand much more of the content of the messages they are handling. They are able to provide services such as publish/subscribe that deliver messages to clients according to their interest in the subject of the message, and message transformation that ensure messages are delivered to applications in the format that they require. These message broker functions remove programming work from the client applications, and permit easier connection of applications that have not been designed to communicate.
Message broking has much in common with federated databases, and the two operate very effectively together. The semantics of messages is very similar to that of relational tuples; the seminar will end with a description of a 'relational message broker'.
Previous musical experiences significantly affect the way people listen to music. Listeners growing up in different cultures can develop vastly different sets of musical preferences, strategies and intuitions.
In recent years, cognitive scientists have been gathering increasing amounts of empirical data which may shed some light on certain aspects of human music cognition. With these findings, various computationally implementable theories have been developed concerning human music cognition and learning.
This talk will highlight some of the results coming out from the MAeStro project, developed to model some of the current theories of music cognition and learning. The system provides insights into what implications these theories have when put into practice. The informationally-rich nature of the musical data provides a challenging environment for machine learning, pattern recognition, and discrete time series prediction approaches. The system is designed according to the multi-agent system paradigm and an information theoretic methodology is used to evaluate the system's performance. A large collection of roughly 8,000 folk songs from different cultures enables highly interesting experiments to be performed.
Many natural language processing tasks require us to be able to automatically translate English sentences into a meaning representation of some kind that we can do further computation with. First and higher order logics are good candidates. But whereas translating sentences like `All men are mortal' into logic is reasonably easy, many sentences contain context-dependent constructs like pronouns (`he') or ellipsis (`is...too'): `All men are mortal; Socrates is a man; therefore he is mortal; Plato is too, since he is also a man.' Until you know the referent of the pronoun, or how the ellipsis is to be completed, you can't do the translation, and knowing these things typically requires inference.
This talk describes an approach to context-dependent translation of such constructs using higher-order unification. This has the benefit that the translation is reversible: we can both translate sentences to logical forms in a way that takes account of contextually dependent elements, and generate from logical forms sentences that relate to their context in an appropriate fashion.
I will give a demonstration of a system implemented along these lines and show how it can be used in conjunction with a theorem prover to solve logical puzzles expressed in English. I will also suggest how it might be used for producing logically consistent English language specifications of system behaviour.
This talk will consider the extent to which emotional responses can be induced in a virtual reality setting. This has been shown to be important in the context of shared virtual environments, where people meet together in order to carry out some joint task. The lack of the ability of avatars to express emotion proved a hindrence to effective communication within small groups. The lessons learned from the series of experimental studies on this led to the development of 'emoting' avatars. These have subsequently been used in an experiment to assess the effect to which speakers to an entirely virtual audience respond with appropriate affect to well-behaved and badly behaved audiences.