Neural Networking

Artificial neural networks (ANNs) are computational paradigms which implement simplified models of their biological counterparts, biological neural networks. Biological neural networks are the local assemblages of neurons and their dendritic connections that form the (human) brain. Accordingly, artificial neural networks are characterised by: Local processing in artificial neurons (processing elements) Massively parallel processing, implemented by rich connection pattern between processing elements The ability to acquire knowledge via learning from experience Knowledge storage in distributed memory, the synaptic processing elements connections The attempt of implementing neural networks for brain like computations like patterns recognition, decisions making, motor control and many others is made possible by the advent of large scale computers in the late 1950's. Indeed artificial neural networks can be viewed as a major new approach to computational methodology since the introduction of digital computers. Although the initial intent of artificial neural networks was to explore and reproduce human information processing tasks such as speech, vision, and knowledge processing, artificial neural networks also demonstrated their superior capability for classification and function approximation problems. This has great potential for solving complex problems such as systems control, data compression, optimization problems, pattern recognition, and system identification. Artificial neural networks were originally developed as tools for the exploration and reproduction of human information processing tasks such as speech, vision, touch, knowledge processing and motor control. Today, most research is directed towards the development of artificial neural networks for applications such as data compression, optimisation, pattern matching, system modeling, function approximation, and control. One of the application areas to which artificial neural networks are applied is flight control. Artificial neural networks give control systems a variety of advanced capabilities. Since artificial neural networks are highly parallel systems, conventional computers are unsuited for neural network algorithms. Special purpose computational hardware has been constructed to efficiently implement artificial neural networks. Accurate Automation has developed a Neural Network Processor. This hardware will allow us to run even the most complex neural networks in real time. The neural network processor is capable of multiprocessor operation in Multiple Instruction Multiple Data (MIMD) fashion. It is the most advanced digital neural network hardware in existence. Each neural network processor system is capable of implementing 8000 neurons with 32,000 interconnections per processor. The computational capability of a single processor 140 million connections per second. An 8 processor neural network processor would be capable of over one billion connections per second. The neural network processor architecture is extremely flexible and any neuron is capable of interconnecting with other neuron in the system. Conventional computers rely on programs that solve a problem using a predetermined series of steps, called algorithms. These programs are controlled by a single, complex central processing unit, and store information at specific locations in memory. Artificial neural networks use highly distributed representations and transformations that operate in parallel, have distributed control through many highly interconnected neurons, and store their information in variable strength connections called synapses. There are many different ways in which people refer to the same type of neural networks technology. Neural networks are described as connectionist systems, because of the connections between individual processing nodes. They are sometimes called adaptive systems, because the values of these connections can change so that the neural network performs more effectively. They are also sometimes called parallel distributed processing systems, which emphasise the way in which the many nodes or neurons in a neural network operate in parallel. The theory that inspires neural network systems is drawn from many disciplines, primarily from neuroscience, engineering, and computer science, but also from psychology, mathematics, physics, and linguistics. These sciences are working toward the common goal of building intelligent systems.

Future of User Interfaces

Several new user interface technologies and interaction principles seem to define a new generation of user interfaces that will move off the flat screen and into the physical world to some extent. Many of these next generation interfaces will not have the user control the computer through commands, but will have the computer adapt the dialogue to the user's needs based on its inferences from observing the user. Most current user interfaces are fairly similar and belong to one of two common types: Either the traditional alphanumeric full screen terminals with a keyboard and function keys, or the more modern WIMP workstations with windows, icons, menus, and a pointing device. In fact, most new user interfaces released after 1983 have been remarkably similar. In contrast, the next generation of user interfaces may move beyond the standard WIMP paradigm to involve elements like virtual realities, head mounted displays, sound and speech, pen and gesture recognition, animation and multimedia, limited artificial intelligence, and highly portable computers with cellular or other wireless communication capabilities. It is hard to envision the use of this hodgepodge of technologies in a single, unified user interface design, and indeed, it may be one of the defining characteristics of the next generation user interfaces that they abandon the principle of conforming to a canonical interface style and instead become more radically tailored to the requirements of individual tasks. The fundamental technological trends leading to the emergence of several experimental and some commercial systems approaching next generation capabilities certainly include the well known phenomena that CPU speed, memory storage capacity, and communications bandwidth all increase exponentially with time, often doubling in as little as two years. In a few years, personal computers will be so powerful that they will be able to support very fancy user interfaces, and these interfaces will also be necessary if we are to extend the use of computers to larger numbers than the mostly penetrated markets of office workers. Traditional user interfaces were function oriented, the user accessed whatever the system could do by specifying functions first and then their arguments. For example, to delete a file in a line-oriented system, the user would first issue the delete command in some way such as typing delete. The user would then further specify that the name of the item to be deleted. The typical syntax for function oriented interfaces was a verb noun syntax. In contrast, modern graphical user interfaces are object oriented, the user first accesses the object of interest and then modifies it by operating upon it. There are several reasons for going with an object oriented interface approach for graphical user interfaces. One is the desire to continuously depict the objects of interest to the user to allow direct manipulation. Icons are good at depicting objects but often poor at depicting actions, leading objects to dominate the visual interface. Furthermore, the object oriented approach implies the use of a noun verb syntax, where the file is deleted by first selecting the file and then issuing the delete command (for example by dragging it into the recycle bin). With this syntax, the computer has knowledge of the operand at the time where the user tries to select the operator, and it can therefore help the user select a function that is appropriate for that object by only showing valid commands in menus and such. This eliminates an entire category of syntax errors due to mismatches between operator and operand. A further functionality access change is likely to occur on a macro level in the move from application oriented to document oriented systems. Traditional operating systems have been based on the notion of applications that were used by the user one at a time. Even window systems and other attempts at application integration typically forced the user to use one application at a time, even though other applications were running in the background. Also, any given document or data file was only operated on by one application at a time. Some systems allow the construction of pipelines connecting multiple applications, but even these systems still basically have the applications act sequentially on the data. The application model is constraining to users who have integrated tasks that require multiple applications to solve. Approaches to alleviate this mismatch in the past have included integrated software and composite editors that could deal with multiple data types in a single document. No single program is likely to satisfy all computer users, however, no matter how tightly integrated it is, so other approaches have also been invented to break the application barrier. Cut and paste mechanisms have been available for several years to allow the inclusion of data from one application in a document belonging to another application. Recent systems even allow live links back to the original application such that changes in the original data can be reflected in the copy in the new document (such as Microsoft`s OLE technology). However, these mechanisms are still constrained by the basic application model that require each document to belong to a specific application at any given time. An alternative model is emerging in object oriented operating systems where the basic object of interest is the user's document. Any given document can contain sub objects of many different types, and the system will take care of activating the appropriate code to display, print, edit, or email these data types as required. The main difference is that the user no longer needs to think in terms of running applications, since the data knows how to integrate the available functionality in the system. In some sense, such an object oriented system is the ultimate composite editor, but the difference compared to traditional, tightly integrated multi-media editors is that the system is open and allows plug and play addition of new or upgraded functionality as the user desires without changing the rest of the system. Even the document oriented systems may not have broken sufficiently with the past to achieve a sufficient match with the users' task requirements. It is possible that the very notion of files and a file system is outdated and should be replaced with a generalised notion of an information space with interlinked information objects in a hypertext manner. As personal computers get multi Gigabyte harddisks, and additional Terabytes become available over the Internet, users will need to access hundreds of thousands or even millions of information objects. To cope with this mass of information, users will need to think of them in more flexible ways than simply as files, and information retrieval facilities need to be made available on several different levels of granularity to allow users to find and manipulate associations between their data. In addition to hypertext and information retrieval, research approaching this next generation data paradigm includes the concept of piles of loosely structured information objects, the information workspace with multiple levels of information storage connected by animated computer graphics to induce a feeling of continuity, personal information management systems where information is organised according to the time it was accessed by the individual user, and the integration of fisheye hierarchical views of an information space with feedback from user queries. Also, several commercial products are already available to add full text search capabilities to existing file systems, but these utility programs are typically not integrated with the general file user interface.

Machine Translation

Serious efforts to develop machine translation systems were under way soon after the ENIAC was built in 1946, and the first known public trial took place in January 1954 at Georgetown University. We've made remarkable progress in the past fifty years, but machine translation involves so many complex tasks that current systems give only a rough idea or indication of the topic and content of the source document. These systems tend to reach a quality barrier beyond which they cannot go, and work best if the subject matter is specific or restricted, free of ambiguities and straightforward, the typical example for this type of text are computer manuals. We'll need more advanced systems to handle the ambiguities and inconsistencies of the real world languages, no wonder that translation, whether performed by machine or by human, is often regarded as an art discipline, and not as an exact science. Today we have an entire range of translation methods with varying degrees of human involvement. Two extreme ends of the translation spectrum are fully automatic high quality translation which has no human involvement and traditional human translation which has no machine translation input. Human aided machine translation and machine aided human translation lie in between these extremities, with human aided machine translation being primarily a machine translation system that requires human aid, whereas machine aided human translation is a human translation method that is utilising machine translation as an aid or tool for translation. The term Computer Assisted Translation is often used to represent both human aided machine translation and machine aided human translation. So, what's so hard about machine translation? There's no such thing as a perfect translation, even when performed by a human expert. A variety of approaches to machine translation exist today, with direct translation being the oldest and most basic one of them. It translates texts by replacing source language words with target language words, with the amount of analysis varying from system to system. Typically it would contain the correspondence lexicon, lists of source language patterns and phrases and mappings to their target language equivalents. The quality of the translated text will vary depending on a size of the system's lexicon and on how smart the replacement strategies are. The main problem with this approach is its lack of contextual accuracy and the inability to capture the real meaning of the source text. Going a step further, syntactic transfer systems use software parsers to analyse the source-language sentences and apply linguistic and lexical rules (or transfer rules) to rewrite the original parse tree so it obeys the syntax of the target language. On the other hand, interlingual systems translate texts using a central data representation notation called an interlingua. This representation is neutral to any languages in the system and breaks the direct relationship that a bilingual dictionary approach would have. Statistical systems use standard methods for translation, but their correspondence lexicon are constructed automatically, using advanced alignment algorithms from a large amount of text for each language, usually available in online databases.