 Purpose |
are an emerging class of computational tools now ready for real-world applications. These algorithms are usually extremely compute intensive; their intrinsic parallelism must be exploited to satisfy the time constraints of many practical applications. The present project aimed at showing the feasibility of a systolic accelerator for some of the most common neural algorithms. An application of Kohonen's Self-Organizing Feature Maps to power-system monitoring was chosen as a reference problem. |
|
|
|
Keyresults |
One of the few existing multi-model neurocomputers has been designed
and build. It is a massively parallel accelerator, called
, based on a bidimensional systolic array with up to
40 x 40 custom processing elements named Genes IV .
Several original features of the array make it possible to avoid some typical inefficiencies of systolic architectures. A system with 20 x 20 processing elements has been demonstrated and used for performance assessment. A complete development system (including the integration in a ) has been built. Novel techniques to simplify the task of writing low-level microcode have been developed; these should be general enough to be ported onto similar systems. The systolic hardware imposes some modifications to the implemented algorithms. For Kohonen's maps the changes are non-negligible and demanded a theoretical verification that the properties of the original are preserved. The modified algorithm is also appropriate other highly pipelined machines whose requirements are similar to Mantra I's. |
|
|
|
|
Methods |
Two custom circuits have been designed: the first for the custom
processing element composing the bidimensional array and the other
for an auxiliary linear array. These circuits have been designed
using VLSI Technology 1,mu m Standard Cell libraries to reduce
design time and to produce fully operational dies in the first place.
Programming the different parts of the machine has been performed with standard tools and languages whenever possible. The techniques to produce efficient code for the systolic array have been adapted from traditional microcode compaction techniques integrated with new ad hoc concepts. Theoretical studies on the modifications to Kohonen's algorithm have been performed with the same analytical tools used for the classic algorithm. |
|
|
|
|
Validation |
Two common algorithms have been implemented on the system: the
and the Kohonen algorithm. These two
implementations have been used to evaluate the sustained
performances of the system. The speed has been found close to peak
performances on large problems, whilst smaller problems are
predictably more critical.
The theoretical studies on the modifications to the Kohonen algorithm have been first backed by extensive simulations on traditional floating-point hardware. The simulations confirmed the substantial similarity of the modified algorithm to the original one for all practical purposes. A large problem of quantisation of a speech database has been used in the simulations. On the machine, where the algorithmic modifications are supplemented by fixed-point arithmetics and other constraints for efficient hardware utilisation, both an original power-system monitoring application and the quantisation problem have been run; the differences with results obtained on traditional hardware have both found very small. The power-system application running on Mantra~I has been demonstrated at the International Conference on Intelligent System Application to Power Systems, held in Montpellier (F) in 1994. The need to compare the performances of Mantra I with those of other systems triggered contacts with many other groups performing similar works in the world and consequently some surveys of the domain have been published. Preliminary performance comparisons have been attempted in cooperation with leading groups from Adaptive Solutions (USA) and Siemens. Although precise comparisons are difficult because of the experimental nature and heterogeneity of the systems, Mantra I appears to have performances in the same range of other research systems. |
|
|
|
|
Perspectives |
Accelerators based on programmable systolic arrays are rare and
currently far from industrialisation. This project made it possible
to identify those problems whose solution is fundamental to make
this kind of massively parallel hardware attractive to potential
users.
In particular, the complexity of producing efficient code is a problem that is severely hindering the commercial launch of machines similar to Mantra~I. A FNRS project has been recently started to address the problem in a general context. The purpose is to provide this kind of accelerator with programming tools similar in comfort to those available in more traditional environments. |
|
|
|
|
Contact |
MANTRA Project Address: EPFL - Mantra Centre INJ - Ecublens 1015 Lausanne Tel: +41-21-693 6619 Fax: +41-21-693 5263 Several people contributed to the project and can be contacted for further information on specific topics:
|
|
|
|
|
Purpose |
The aim of the project is to build an automatic system capable of recognizing the cloud amounts (during the night), from data collected by various meteorological instruments including pyrgeometers and ceilometers. This system is to be part of a global automatic METAR generator, developed at the ISM, in the project Ametis~II. More precisely, the system must detect the number of cloud layers, and for each layer it must give its height and cloudiness. |
|
|
|
|
Keyresults |
The first results obtained were on the recognition of the total
cloudiness, which is a meteorological value of interest. The neural
network performs a recognition rate of 90% on the training set and
88% on a test set of unlearned examples, with 9 hidden neurons.
The whole system detects in addition the heights of several cloud layers and estimates the cloud amount of each one separately. It performs a recognition rate of 81% on a one-year test set. |
|
|
|
|
Methods |
The heart of the system is a feedforward multilayer neural network,
trained to recognize the cloud amount of a cloud layer. We developed
and used a quasi-Newtonian second order optimization method to perform
training. Special attention was devoted to the preprocessing of the
large amount of heterogeneous data available.
The rest of the system consists in processings by rules to detect the number and heights of the cloud layers, and to produce a METAR compliant output. |
|
|
|
|
Validation |
The validation of the Clouds module is in progress, as the development of the Ametis~II system. It will be tested at Zurich airport from summer 1996 while the whole system will be completed. In 1997 the fully automated observations will be operationally displayed. |
|
|
|
|
Perspectives |
The next step will be the development of an automated observation system at Geneva airport (1998-2001). According to the success of this transfer of technology to another airport, the AMETIS 2 concept could be generalized to other secondary airports in Switzerland, after year 2000. |
|
|
|
|
Contact |
MANTRA Project Address: EPFL - Mantra Centre INJ - Ecublens 1015 Lausanne Tel: +41-21-693 6619 Fax: +41-21-693 5263
Thierry Cornu
Frédéric Aviolat
Daniel Cattani |
|
|
|
|
Partners |
This project is part of the Ametis II project of the Swiss Institute of Meteorology (ISM), which aims at generating automatically complete METAR messages. The MANTRA-B project designed the ``Clouds module'' of this global system. Collaboration with Technikum Rapperswil took place for the integration of our module into the Ametis II system. All data was provided by the ISM. |
|
|
|
|
Purpose |
The purpose of this project is to investigate the possibility of improving speech recognition, using an analog VLSI artificial cochlea. Key issues are the identification of the processing performed by the human cochlea, the reproduction of this processing with an artificial cochlea, and the interfacing of the artificial cochlea with a standart speech recognition system. |
|
|
|
|
Keyresults |
In the scope of this project, an analog VLSI cochlea has been integrated and tested. This artificial cochlea has been presented at the NIPS'95 conference, and is now the analog VLSI cochlea with the most regular filter characteristics published to date. Based on this cochlea, a linear predictive coding (LPC) scheme has been developed, which replaces the LPC delay line with our artificial cochlea, yielding a Cochlear Linear Predictive Coding (CLPC). The CLPC pre-processing has been used as front-end for a traditional speech recognition system, that uses vector quantization and Hidden Markov Models. It has been shown that the CLPC pre-processing outperforms the standart LPC processing, but underlies cepstral coding, which is today's most used pre-processing. |
|
|
|
|
Methods |
The artificial cochlea has been designed using the Compass
layout tools, and has been integrated in ES2's 1.5 micron
technology on a 4.8mm x 3.2mm die. In order to simulate the Cochlear Linear Predictive Coding a C program has been written which runs on a Sun Sparc 20 WorkStation. |
|
|
|
|
Validation |
Validation of the CLPC approach has been obtained in collaboration with the CIRC laboratory at the EPFL in the framework of the European Esprit project HIMARNNET. The CLPC simulation has been used to pre-process the HIMARNNET multi-speaker, isolated word, telephone quality database, which contained about 50 words pronounced by 280 speakers. The resulting data has been parsed by a Vector Quantizer and Hidden Markov Models developed for the HIMARNNET project. |
|
|
|
|
Perspectives |
We will continue our research in this domain in the framework of project founded by the Swiss National Fund for Scientific Research. This project will address speech recognition in a noisy environment and features extracted at a higher level in the human auditory pathway will be used to seperate the speaker signal from the background noise. In this project we will implement more of the human auditory pathway in analog VLSI, before interfacing with a classical recognition system. |
|
|
|
|
Contact |
MANTRA Project EPFL - Mantra Centre INJ - Ecublens 1015 Lausanne Tel: +41-21-693 6619 Fax: +41-21-693 5263 Eric Fragnière E-mail: Eric.Fragniere@di.epfl.ch |
|
|
|
|
Partners |
|
|
|
|