Neural networking – attempts by scientists and engineers in the US and Japan to build a computer brain which simulates the learning process in human beings – is now becoming a subject for serious research amongst Europe’s artificial intelligentsia. Brain, or Basic Research in Adaptive Intelligence, is an EEC funded pan-European research project involving engineers from Siemens, Thomson, Philips and Plessey, while in June this year, Paris will play host to Europe’s first neural networking conference, nEuro ’88. The conference will be co-ordinated by the Ecole Superieure de Physique et de Chemie Industrielle, and the Ecole Superieure des Telecommunications, two of the leading bodies within the French neural field. Representatives from the so-called Grandes Ecoles will also undoubtedly join a working group shortly to be set up at the bidding of the French high-tech watchdog, OFTA, Observatoire Franc,ais des Techniques Avance’es. The group plans to unite experts from fields as diverse as neurobiology and mathematics to debate neural networking issues, and will kick off by considering the effect of synthesis on moving neural technology forward, an appropriate training programme for students, and the potential for co-operation between different research teams. Recently, the Ecole Superieure de Physique joined forces with the Ecole Polytechnique to build an integrated circuit for use within a concurrent fault tolerant project. Scientists working on the circuit will be drawing upon past experiments which show that, by fixing rows of active elements with individual memory on to to a board, it is possible to simulate memory cells in the human brain. The elements, or neurons, are encouraged to interact and forge synaptic connections with each other, and once a stable environment has been established, the memory can be taught to recognise and re enact a task. For example, the noise of the spoken word brain can be entered into the memory synapses, which then interconnect until they identify the written word brain, and place it on the output screen. Neural networks such as Net Talk have used a similar process to learn to speak English; Net Talk has even managed to deduce appropriate irregularities of pronunciation in new words, by drawing upon the implicit rules that govern the words memorised in its vocabulary. Scientists in the Ecole de Physique’s neural network department are hoping to integrate the fruits of these experiments onto a single 64 neuron-capacity chip, and aim long-term to implant them in machines designed with military, industrial and space applications in mind. The Ecole de Physique is also taking part in a similar project initiated by the French Industry and Research ministry which aims to build a Transputer-based neuro computer. Optical neural-computing, where electrical switching elements are replaced by optical ones, with resulting gains in speed and a noticeable reduction of isolation problems, is also on the French neural agenda. A project has been set up by the Jessy de Toulon and the Ecole de Physique to build a neural computer that will run eight times as fast as 68020-based Apollo Computer Domain 3000 workstations. If they succeed, France’s place on the neural map will be assured; so far, the world’s two optical neural computers have been developed by Mitsubishi in Japan, and researchers at the California Institute of Technology in Berkeley, US.
Content from our partners
