Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
  • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means

Definitions

• the present invention relates to a weighting network, preferably for use in a neural network.
• So-called artificial neural networks are preferably used in the complex processing of several input variables to determine a solution.
• Special fields of application are image processing, speech processing and signal processing in general, such as the processing of data for material testing and the processing of medical data.
• Conventional weighting networks consist of a plurality of processing devices connected in parallel, each of which has a plurality of weighting devices connected in parallel.
• a number of input signals are fed to each of the processing devices, in each of which an input signal is fed to a weighting device.
• the individual input signal weighted, ie multiplied by a certain factor or changed with any function.
• the output signals of the weighting devices are fed together to a summing device present in the associated processing device and added up.
• the output signals of the summing devices are either applied to a subsequent weighting network of the same type or by feedback to the input of the same weighting network.
• a pattern comparison takes place, in which the various output signals of the summing units are compared with comparison patterns and the comparison pattern of greatest similarity is selected and output.
• a weighting network is described in the article Neurocomputer in practical use by V. David shenchez A., CH-Z: Technische Rundschau, 25/90, pages 60-65.
• the present invention is based on the object of providing a weighting network, in particular for a neural network, which enables improved and reliable solution finding.
• the present invention is based on the basic idea of creating a competitive situation within the weighting network.
• the output signals of the summing devices are evaluated in a comparator device and, depending on the result, forwarded with a different time delay for the next processing.
• the output signals are evaluated using comparison criteria, in which the greatest similarity to a known variable is preferably determined or a comparison of the output signals with one another is used.
• weighting networks of the present invention can be connected in series, for example in an input layer, several intermediate layers and an output layer of a neural network.
• a certain output signal with the least time delay can be passed to the first intermediate layer and when passing through the first intermediate layer if the resulting output signals of the summing devices have only a low similarity value have, with a large delay.
• the signals supplied by the output signals of the summing device of a weighting network, the similarity value of which is below a certain threshold value, are blocked, ie are not permitted for further processing in a subsequent intermediate layer.
• Fig. 1 shows a first embodiment of an inventive
• Fig. 2 shows a second embodiment of an inventive
• FIG. 3 is a timing diagram of signals as they may arise at the output of the embodiment of FIG. 2.
• the embodiment of a weighting network according to the invention shown in FIG. 1 receives a number (i) of input signals el, e2, ..., el, ... ei.
• Each of the input signals is processed in an associated weighting device Gll, G12, ..., Gll, ... Gli with a preferably adjustable weighting factor or a weighting function.
• All output signals of the weighting devices are fed to a summing device S1.
• a fixedly predetermined input signal (+1) can be passed through a weighting device Glz and its output signal can also be fed to the summing device S1.
• the output signal of the summing device S1 is fed to a downstream delay device D1 with a variable, adjustable delay.
• the delay set depends on the value of the output signal of the summing device which is fed to the delay device.
• a comparator device C1 is provided, which receives the output signal of the summing device S1 on the one hand and a reference signal R on the other hand and by a Comparison produces a signal with which the effective time delay of the delay device D1 can be set.
• FIG. 1 shows a weighting network with a processing device VI.
• a plurality of such processing devices can be arranged parallel to one another, the input signals el, e2, ... el, ... ei also being applied to the further processing devices.
• a plurality of weighting networks or groups are connected in series and form the input layer, a plurality of intermediate layers and the output layer of a neural network.
• the weighting network shown has several processing devices VI; V2 ..., Vk, ... Vm, which are connected in parallel to each other and receive input signals el, e2, ... el, ... ei.
• Each processing device has several weighting devices Gll, G12, ... Gll, ... Gli; ...; G l, Gm2, ... Gml, ..., G i on.
• the output signals are summed up in summing devices S1, S2, ..., Sk, ..., Sm.
• the output signals of the summing devices are fed to a comparator device C.
• This comparing device compares the supplied output signals of the summing devices with one another and sorts them on the basis of their values.
• the comparator device C also has a timing control circuit with which the output of the supplied output signals is controlled in such a way that the first-ranked signal after the sorting is output first and the other output signals are output in succession according to their sorting.
• ! _. 2 shows an example of a time diagram, which shows the chronological order in which signals S1, S2, -6 are output. As shown, the signal S6 is first output without a time delay by the comparator device C for subsequent processing. Next follow the signals S ' 4, Sl, S5, S2 and S3.
• the signals output by the comparison device C can, as shown in FIG. 2, be fed to a further weighting network, alternatively a return to the same weighting network is possible.
• the weighting network according to the invention represents a novel type of processing of input signals, with which input signals and the resulting output signals are to a certain extent in a competitive situation.
• the output signals of a first layer are made available for further processing in a further layer in accordance with the value of the individual signals with a corresponding chronological sequence.
• one or more of the best-rated output signals can only be output on their own, i.e. the remaining signals are not permitted for further processing by a blocking device (not shown).
• the present invention has the significant advantage that during processing, in particular in a multilayer neural network, certain signals arrive for further processing in a controlled manner.
• Similarity criteria are sufficient. In this case, less relevant signals are preferably blocked.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• Health & Medical Sciences (AREA)
• Life Sciences & Earth Sciences (AREA)
• Biomedical Technology (AREA)
• Biophysics (AREA)
• Theoretical Computer Science (AREA)
• Evolutionary Computation (AREA)
• General Engineering & Computer Science (AREA)
• Data Mining & Analysis (AREA)
• Artificial Intelligence (AREA)
• General Health & Medical Sciences (AREA)
• Molecular Biology (AREA)
• Computing Systems (AREA)
• Computational Linguistics (AREA)
• General Physics & Mathematics (AREA)
• Mathematical Physics (AREA)
• Software Systems (AREA)
• Neurology (AREA)
• Image Analysis (AREA)
• Feedback Control In General (AREA)

Applications Claiming Priority (2)

Publications (1)

Family

ID=6425713

Family Applications (1)

Country Status (3)

Families Citing this family (3)

Family Cites Families (1)

• 1991
  • 1991-02-22 DE DE4105669A patent/DE4105669A1/de active Granted
• 1992
  • 1992-02-17 EP EP92904977A patent/EP0583245A1/de not_active Withdrawn
  • 1992-02-17 WO PCT/EP1992/000339 patent/WO1992015072A1/de not_active Application Discontinuation

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events