TOP > 外国特許検索 > LEARNING METHOD AND LEARNING DEVICE EMPLOYING AUGMENTATION

LEARNING METHOD AND LEARNING DEVICE EMPLOYING AUGMENTATION

Foreign code F210010489
File No. (S2020-0058-N0)
Posted date 2021年7月29日
Country 世界知的所有権機関(WIPO)
International application number 2020JP043248
International publication number WO 2021100818
Date of international filing 令和2年11月19日(2020.11.19)
Date of international publication 令和3年5月27日(2021.5.27)
Priority data
  • 特願2019-209179 (2019.11.19) JP
Title LEARNING METHOD AND LEARNING DEVICE EMPLOYING AUGMENTATION
Abstract In relation to machine learning, the objective of the present invention is to provide a learning method and device with which it is possible to improve generalization capability and to improve recognition accuracy. A plurality of items of learning data after augmentation with respect to multidimensional quantity learning data before augmentation are input into one classifier, and from among probability distributions of prediction labels output for each of the plurality of items of learning data, at least one probability distribution, selected using the error compared with the correct answer of the learning data before augmentation as a scale, is used to perform learning on the basis of the error compared with the correct answer of the learning data before augmentation. Data sampled with the probability of selection increased in accordance with the magnitude of the error compared with the correct answer, data with which the error compared with the correct answer is greatest, or data with which the difference from the average value of the error compared with the correct answer is smallest are used as the data selected using the error compared with the correct answer as a scale.
Outline of related art and contending technology BACKGROUND ART
In recent years, AI (Artificial Intelligence) has made an outstanding evolution, such as due to technological development of calculators. A central technique is a neural network (NN ; Neural Network), and in particular, a DNN deep in neural network is seating the technique. In general, since DNN's have deeper strata, many research has been conducted on designing deep strata DNN and how well learns deep strata DNN.
However, in order to make the tier deeper, a large amount of data is required for learning, but there is a problem in that the cost of preparing the data is high in the real world. Further, the problem caused by deeper layers is that over-learning is likely to occur. This is to fit the data too much because the model is too highly representative. Furthermore, in a case that the depth of the layer increases the number of parameters, the model size increases, making it difficult to use in a mobile terminal, and the computational complexity at the time of prediction increases.
On the other hand, the classifier of DNN is required to have generalized performance. In other words, it is necessary to be able to accurately predict even unknown data. Therefore, error calculation (cross-entropy) is performed between the probability distribution of the prediction label and the true probability distribution, and parameters are optimized so as to reduce errors. As a device for improving prediction accuracy, during training, data at hand is generally increased in volume (data augmentation) by processing such as scaling, inverting, rotating, or adjusting contrast to increase the data. Also known is an ensemble learning technique in which a plurality of predictors are learned and these prediction results are integrated during operation.
In addition, examples of techniques for ensembling during learning include techniques relating to convolutional neural networks for image recognition focusing on ensemble learning (see NpL 1); A technique related to a neural network having a residual structure (see Non-Patent Document 2) and a learning processing method for identifying whether or not a disease is present on the basis of skin image data (see Patent Document 1) are disclosed. However, any of the above-described documents utilizes the output results predicted by the plurality of learning apparatuses, and does not utilize the output results obtained by causing the plurality of pieces of learning data after the addition of water to be input to one learning apparatus. In addition, Non-PTL 1 and PTL 1 disclose technology in which learning data is pooled by rotating and reversing an image, but among a plurality of pieces of learning data after the pooled learning data, the plurality of pieces of learning data include, by being used for learning, Data that degrades recognition accuracy may exist, and as a result, there is a problem in that recognition accuracy cannot be sufficiently improved when the data is used as the learning data after the addition of water.
  • Applicant
  • ※All designated countries except for US in the data before July 2012
  • KWANSEI GAKUIN EDUCATIONAL FOUNDATION
  • Inventor
  • OKADOME Takeshi
  • IDE Atsuya
IPC(International Patent Classification)
Specified countries National States: AE AG AL AM AO AT AU AZ BA BB BG BH BN BR BW BY BZ CA CH CL CN CO CR CU CZ DE DJ DK DM DO DZ EC EE EG ES FI GB GD GE GH GM GT HN HR HU ID IL IN IR IS IT JO JP KE KG KH KN KP KR KW KZ LA LC LK LR LS LU LY MA MD ME MG MK MN MW MX MY MZ NA NG NI NO NZ OM PA PE PG PH PL PT QA RO RS RU RW SA SC SD SE SG SK SL ST SV SY TH TJ TM TN TR TT TZ UA UG US UZ VC VN WS ZA ZM ZW
ARIPO: BW GH GM KE LR LS MW MZ NA RW SD SL SZ TZ UG ZM ZW
EAPO: AM AZ BY KG KZ RU TJ TM
EPO: AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR
OAPI: BF BJ CF CG CI CM GA GN GQ GW KM ML MR NE SN ST TD TG
ライセンスをご希望の方、特許の内容に興味を持たれた方は、下記「問合せ先」までお問い合わせください。

PAGE TOP

close
close
close
close
close
close