Construction Of Training Set Research Articles

Unsupervised Seismic Random Noise Suppression Based on Local Similarity and Replacement Strategy

Improving the signal-to-noise ratio and suppressing random noise in seismic data is critical for high-precision processing. Although deep learning-based algorithms have gained popularity as denoising methods, they suffer from poor generalization ability, resulting in high training set construction cost and computation cost. To address this problem, we propose an unsupervised learning-based denoising method that includes an improved denoising strategy based on local similarity and replacement, a corresponding training method, and an improved network based on UNet. Our training method takes advantage of network convergence and allows direct training on the test region, effectively solving the problems associated with denoising methods using generalization ability while improving training performance. In addition, our network is specifically designed for the training method and incorporates various improvements that could further enhance the training effectiveness. Our method outperforms traditional denoising methods, as demonstrated by tests on synthetic and field data, with superior performance in random noise attenuation and reflection event reconstruction.

Training Set Construction for Genomic Prediction in Auto-Tetraploids: An Example in Potato.

Training set construction is an important prerequisite to Genomic Prediction (GP), and while this has been studied in diploids, polyploids have not received the same attention. Polyploidy is a common feature in many crop plants, like for example banana and blueberry, but also potato which is the third most important crop in the world in terms of food consumption, after rice and wheat. The aim of this study was to investigate the impact of different training set construction methods using a publicly available diversity panel of tetraploid potatoes. Four methods of training set construction were compared: simple random sampling, stratified random sampling, genetic distance sampling and sampling based on the coefficient of determination (CDmean). For stratified random sampling, population structure analyses were carried out in order to define sub-populations, but since sub-populations accounted for only 16.6% of genetic variation, there were negligible differences between stratified and simple random sampling. For genetic distance sampling, four genetic distance measures were compared and though they performed similarly, Euclidean distance was the most consistent. In the majority of cases the CDmean method was the best sampling method, and compared to simple random sampling gave improvements of 4–14% in cross-validation scenarios, and 2–8% in scenarios with an independent test set, while genetic distance sampling gave improvements of 5.5–10.5% and 0.4–4.5%. No interaction was found between sampling method and the statistical model for the traits analyzed.

Presentation Attack Detection on Limited-Resource Devices Using Deep Neural Classifiers Trained on Consistent Spectrogram Fragments.

The presented paper is concerned with detection of presentation attacks against unsupervised remote biometric speaker verification, using a well-known challenge–response scheme. We propose a novel approach to convolutional phoneme classifier training, which ensures high phoneme recognition accuracy even for significantly simplified network architectures, thus enabling efficient utterance verification on resource-limited hardware, such as mobile phones or embedded devices. We consider Deep Convolutional Neural Networks operating on windows of speech Mel-Spectrograms as a means for phoneme recognition, and we show that one can boost the performance of highly simplified neural architectures by modifying the principle underlying training set construction. Instead of generating training examples by slicing spectrograms using a sliding window, as it is commonly done, we propose to maximize the consistency of phoneme-related spectrogram structures that are to be learned, by choosing only spectrogram chunks from the central regions of phoneme articulation intervals. This approach enables better utilization of the limited capacity of the considered simplified networks, as it significantly reduces a within-class data scatter. We show that neural architectures comprising as few as dozens of thousands parameters can successfully—with accuracy of up to 76%, solve the 39-phoneme recognition task (we use the English language TIMIT database for experimental verification of the method). We also show that ensembling of simple classifiers, using a basic bagging method, boosts the recognition accuracy by another 2–3%, offering Phoneme Error Rates at the level of 23%, which approaches the accuracy of the state-of-the-art deep neural architectures that are one to two orders of magnitude more complex than the proposed solution. This, in turn, enables executing reliable presentation attack detection, based on just few-syllable long challenges on highly resource-limited computing hardware.

Supervised-learning link prediction in single layer and multiplex networks

The emergence of complex real-world networks has put forth a plethora of information about different domains. Link-prediction is one of the emerging research problems that utilizes the information from the networks to find future relationships between the nodes. The structure of real-world networks varies from having homogeneous relationships to having multiple associations. The homogeneous relationships are modeled by single-layer networks, while the multiplex networks represent the multiple associations. This study proposes a solution for finding future links in single-layer and multiplex networks by using supervised machine learning techniques. This study considers a set of topological features of the network for training the machine learning classifiers. The training and testing data set construction framework devised in this work helps in evaluating the proposed method on different networks. This study also contributes towards identifying four community-based features for the proposed mechanism.

Fraud Detection in Batches of Sweet Almonds by Portable Near-Infrared Spectral Devices

One of the key challenges for the almond industry is how to detect the presence of bitter almonds in commercial batches of sweet almonds. The main aim of this research is to assess the potential of near-infrared spectroscopy (NIRS) by means of using portable instruments in the industry to detect batches of sweet almonds which have been adulterated with bitter almonds. To achieve this, sweet almonds and non-sweet almonds (bitter almonds and mixtures of sweet almonds with different percentages (from 5% to 20%) of bitter almonds) were analysed using a new generation of portable spectrophotometers. Three strategies (only bitter almonds, bitter almonds and mixtures, and only mixtures) were used to optimise the construction of the non-sweet almond training set. Models developed using partial least squares-discriminant analysis (PLS-DA) correctly classified 86–100% of samples, depending on the instrument used and the strategy followed for constructing the non-sweet almond training set. These results confirm that NIR spectroscopy provides a reliable, accurate method for detecting the presence of bitter almonds in batches of sweet almonds, with up to 5% adulteration levels (lower levels should be tested in future studies), and that this technology can be readily used at the main steps of the production chain.

Gaia Early Data Release 3

Aims. We produce a clean and well-characterised catalogue of objects within 100 pc of the Sun from the Gaia Early Data Release 3. We characterise the catalogue through comparisons to the full data release, external catalogues, and simulations. We carry out a first analysis of the science that is possible with this sample to demonstrate its potential and best practices for its use. Methods. Theselection of objects within 100 pc from the full catalogue used selected training sets, machine-learning procedures, astrometric quantities, and solution quality indicators to determine a probability that the astrometric solution is reliable. The training set construction exploited the astrometric data, quality flags, and external photometry. For all candidates we calculated distance posterior probability densities using Bayesian procedures and mock catalogues to define priors. Any object with reliable astrometry and a non-zero probability of being within 100 pc is included in the catalogue. Results. We have produced a catalogue of 331 312 objects that we estimate contains at least 92% of stars of stellar type M9 within 100 pc of the Sun. We estimate that 9% of the stars in this catalogue probably lie outside 100 pc, but when the distance probability function is used, a correct treatment of this contamination is possible. We produced luminosity functions with a high signal-to-noise ratio for the main-sequence stars, giants, and white dwarfs. We examined in detail the Hyades cluster, the white dwarf population, and wide-binary systems and produced candidate lists for all three samples. We detected local manifestations of several streams, superclusters, and halo objects, in which we identified 12 members of Gaia Enceladus. We present the first direct parallaxes of five objects in multiple systems within 10 pc of the Sun. Conclusions. We provide the community with a large, well-characterised catalogue of objects in the solar neighbourhood. This is a primary benchmark for measuring and understanding fundamental parameters and descriptive functions in astronomy.

A neural network-based methodology to select young stellar object candidates from IR surveys

Context. Observed young stellar objects (YSOs) are used to study star formation and characterize star-forming regions. For this purpose, YSO candidate catalogs are compiled from various surveys, especially in the infrared (IR), and simple selection schemes in color-magnitude diagrams (CMDs) are often used to identify and classify YSOs. Aims. We propose a methodology for YSO classification through machine learning (ML) using Spitzer IR data. We detail our approach in order to ensure reproducibility and provide an in-depth example on how to efficiently apply ML to an astrophysical classification. Methods. We used feedforward artificial neural networks (ANNs) that use the four IRAC bands (3.6, 4.5, 5.8, and 8 μm) and the 24 μm MIPS band from Spitzer to classify point source objects into CI and CII YSO candidates or as contaminants. We focused on nearby (≲1 kpc) star-forming regions including Orion and NGC 2264, and assessed the generalization capacity of our network from one region to another. Results. We found that ANNs can be efficiently applied to YSO classification with a contained number of neurons (∼25). Knowledge gathered on one star-forming region has shown to be partly efficient for prediction in new regions. The best generalization capacity was achieved using a combination of several star-forming regions to train the network. Carefully rebalancing the training proportions was necessary to achieve good results. We observed that the predicted YSOs are mainly contaminated by under-constrained rare subclasses like Shocks and polycyclic aromatic hydrocarbons (PAHs), or by the vastly dominant other kinds of stars (mostly on the main sequence). We achieved above 90% and 97% recovery rate for CI and CII YSOs, respectively, with a precision above 80% and 90% for our most general results. We took advantage of the great flexibility of ANNs to define, for each object, an effective membership probability to each output class. Using a threshold in this probability was found to efficiently improve the classification results at a reasonable cost of object exclusion. With this additional selection, we reached 90% and 97% precision on CI and CII YSOs, respectively, for more than half of them. Our catalog of YSO candidates in Orion (365 CI, 2381 CII) and NGC 2264 (101 CI, 469 CII) predicted by our final ANN, along with the class membership probability for each object, is publicly available at the CDS. Conclusions. Compared to usual CMD selection schemes, ANNs provide a possibility to quantitatively study the properties and quality of the classification. Although some further improvement may be achieved by using more powerful ML methods, we established that the result quality depends mostly on the training set construction. Improvements in YSO identification with IR surveys using ML would require larger and more reliable training catalogs, either by taking advantage of current and future surveys from various facilities like VLA, ALMA, or Chandra, or by synthesizing such catalogs from simulations.

KAT4IA: K-Means Assisted Training for Image Analysis of Field-Grown Plant Phenotypes.

High-throughput phenotyping enables the efficient collection of plant trait data at scale. One example involves using imaging systems over key phases of a crop growing season. Although the resulting images provide rich data for statistical analyses of plant phenotypes, image processing for trait extraction is required as a prerequisite. Current methods for trait extraction are mainly based on supervised learning with human labeled data or semisupervised learning with a mixture of human labeled data and unsupervised data. Unfortunately, preparing a sufficiently large training data is both time and labor-intensive. We describe a self-supervised pipeline (KAT4IA) that uses K-means clustering on greenhouse images to construct training data for extracting and analyzing plant traits from an image-based field phenotyping system. The KAT4IA pipeline includes these main steps: self-supervised training set construction, plant segmentation from images of field-grown plants, automatic separation of target plants, calculation of plant traits, and functional curve fitting of the extracted traits. To deal with the challenge of separating target plants from noisy backgrounds in field images, we describe a novel approach using row-cuts and column-cuts on images segmented by transform domain neural network learning, which utilizes plant pixels identified from greenhouse images to train a segmentation model for field images. This approach is efficient and does not require human intervention. Our results show that KAT4IA is able to accurately extract plant pixels and estimate plant heights.

Atomic Partitioning of the MPn (n = 2, 3, 4) Dynamic Electron Correlation Energy by the Interacting Quantum Atoms Method: A Fast and Accurate Electrostatic Potential Integral Approach.

Recently, the quantum topological energy partitioning method called interacting quantum atoms (IQA) has been extended to MPn (n = 2, 3, 4) wave functions. This enables the extraction of chemical insight related to dynamic electron correlation. The large computational expense of the IQA‐MPn approach is compensated by the advantages that IQA offers compared to older nontopological energy decomposition schemes. This expense is problematic in the construction of a machine learning training set to create kriging models for topological atoms. However, the algorithm presented here markedly accelerates the calculation of atomically partitioned electron correlation energies. Then again, the algorithm cannot calculate pairwise interatomic energies because it applies analytical integrals over whole space (rather than over atomic volumes). However, these pairwise energies are not needed in the quantum topological force field FFLUX, which only uses the energy of an atom interacting with all remaining atoms of the system that it is part of. Thus, it is now feasible to generate accurate and sizeable training sets at MPn level of theory. © 2019 The Authors. Journal of Computational Chemistry published by Wiley Periodicals, Inc.

Computational methods for recognition of cancer protein markers in saliva.

In recent years, many studies have supported that cancer tissues can make disease-specific changes in some salivary proteins through some mediators in the pathogenesis of systemic diseases. These salivary proteins have the potential to become cancer-specific biomarkers in the early diagnosis stage. How to effectively identify these potential markers is one of the challenging issues. In this paper, we propose novel machine learning methods for recognition cancer biomarkers in saliva by two stages. In the first stage, salivary secreted proteins are recognized which are considered as candidate biomarkers of cancers. We picked up 557 salivary secretory proteins from 20379 human proteins by public databases and published literatures. Then, we present a training set construction strategy to solve the imbalance problem in order to make the classification methods get better accuracy. From all human protein set, the proteins belonging to the same families as salivary secretory proteins are removed. After that, we use SVC-KM method to cluster the remaining proteins, and select negative samples from each cluster in proportion. Next, the features of proteins are calculated by tools. We collect 24 protein properties such as sequence, structure and physicochemical properties, a total of 1087 features. An innovative procedure based on the local samples is proposed for selecting the appropriate features, in order to further improve the performance of SVM classifier. Experimental results show that the average sensitivity, specificity and accuracy of salivary secretory protein recognition using selected 32 features in training set are 97.09%, 98.10%, 97.61%, respectively. The use of these methods can improve the accuracy of recognition by solving the problems of unbalanced sample size and uneven distribution in training set. In the second stage, we apply the best model to dig out the salivary secreted proteins from 58 reported cancer markers, and get a total of 42 proteins which are considered to be used for salivary diagnosis. We analyze the gene expression data of three types of cancer, and predict that 33 genes will appear in saliva after they are translated into proteins. This study provides an important computational tool to help biologists and researchers reduce the number of candidate proteins and the cost of research. So as to further accelerate the discovery of cancer biomarkers in saliva and promote the development of saliva diagnosis.

Inference of a compact representation of sensor fingerprint for source camera identification

Sensor pattern noise (SPN) is an inherent fingerprint of imaging devices, which provides an effective way for source camera identification (SCI). Although SPNs extracted from large image blocks usually yield high identification accuracy, their high dimensionality would incur a high computational cost in the matching stage, consequently hindering many applications that require efficient camera matchings. In this work, we employ and evaluate the concept of principal component analysis (PCA) de-noising in SCI tasks. Based on this concept, we present a framework that formulates a compact SPN representation. To enhance the de-noising effect, we introduce a training set construction procedure that minimizes the impact of various interfering artifacts, which is especially useful in some challenging cases, e.g., when only textured reference images are available. To further boost the SCI performance, a novel approach based on linear discriminant analysis (LDA) is adopted to extract more discriminant SPN features. To evaluate our methods, extensive experiments are conducted on the Dresden image database. The results indicate that the proposed framework can serve as an effective post-processing procedure, which not only boosts the performance, but also greatly reduces the computational cost in the matching phase.

Analysis of training sample selection strategies for regression-based quantitative landslide susceptibility mapping methods

All of the quantitative landslide susceptibility mapping (QLSM) methods requires two basic data types, namely, landslide inventory and factors that influence landslide occurrence (landslide influencing factors, LIF). Depending on type of landslides, nature of triggers and LIF, accuracy of the QLSM methods differs. Moreover, how to balance the number of 0 (nonoccurrence) and 1 (occurrence) in the training set obtained from the landslide inventory and how to select which one of the 1's and 0's to be included in QLSM models play critical role in the accuracy of the QLSM. Although performance of various QLSM methods is largely investigated in the literature, the challenge of training set construction is not adequately investigated for the QLSM methods. In order to tackle this challenge, in this study three different training set selection strategies along with the original data set is used for testing the performance of three different regression methods namely Logistic Regression (LR), Bayesian Logistic Regression (BLR) and Fuzzy Logistic Regression (FLR). The first sampling strategy is proportional random sampling (PRS), which takes into account a weighted selection of landslide occurrences in the sample set. The second method, namely non-selective nearby sampling (NNS), includes randomly selected sites and their surrounding neighboring points at certain preselected distances to include the impact of clustering. Selective nearby sampling (SNS) is the third method, which concentrates on the group of 1's and their surrounding neighborhood. A randomly selected group of landslide sites and their neighborhood are considered in the analyses similar to NNS parameters. It is found that LR-PRS, FLR-PRS and BLR-Whole Data set-ups, with order, yield the best fits among the other alternatives. The results indicate that in QLSM based on regression models, avoidance of spatial correlation in the data set is critical for the model's performance.

A comprehensive software suite for protein family construction and functional site prediction.

In functionally diverse protein families, conservation in short signature regions may outperform full-length sequence comparisons for identifying proteins that belong to a subgroup within which one specific aspect of their function is conserved. The SIMBAL workflow (Sites Inferred by Metabolic Background Assertion Labeling) is a data-mining procedure for finding such signature regions. It begins by using clues from genomic context, such as co-occurrence or conserved gene neighborhoods, to build a useful training set from a large number of uncharacterized but mutually homologous proteins. When training set construction is successful, the YES partition is enriched in proteins that share function with the user’s query sequence, while the NO partition is depleted. A selected query sequence is then mined for short signature regions whose closest matches overwhelmingly favor proteins from the YES partition. High-scoring signature regions typically contain key residues critical to functional specificity, so proteins with the highest sequence similarity across these regions tend to share the same function. The SIMBAL algorithm was described previously, but significant manual effort, expertise, and a supporting software infrastructure were required to prepare the requisite training sets. Here, we describe a new, distributable software suite that speeds up and simplifies the process for using SIMBAL, most notably by providing tools that automate training set construction. These tools have broad utility for comparative genomics, allowing for flexible collection of proteins or protein domains based on genomic context as well as homology, a capability that can greatly assist in protein family construction. Armed with this new software suite, SIMBAL can serve as a fast and powerful in silico alternative to direct experimentation for characterizing proteins and their functional interactions.

Efficiency of the spectral-spatial classification of hyperspectral imaging data

The efficiency of methods of the spectral-spatial classification of similarly looking types of vegetation on the basis of hyperspectral data of remote sensing of the Earth, which take into account local neighborhoods of analyzed image pixels, is experimentally studied. Algorithms that involve spatial pre-processing of the raw data and post-processing of pixel-based spectral classification maps are considered. Results obtained both for a large-size hyperspectral image and for its test fragment with different methods of training set construction are reported. The classification accuracy in all cases is estimated through comparisons of ground-truth data and classification maps formed by using the compared methods. The reasons for the differences in these estimates are discussed.

Improvement of Predictive Ability by Uniform Coverage of the Target Genetic Space.

Genome-enabled prediction provides breeders with the means to increase the number of genotypes that can be evaluated for selection. One of the major challenges in genome-enabled prediction is how to construct a training set of genotypes from a calibration set that represents the target population of genotypes, where the calibration set is composed of a training and validation set. A random sampling protocol of genotypes from the calibration set will lead to low quality coverage of the total genetic space by the training set when the calibration set contains population structure. As a consequence, predictive ability will be affected negatively, because some parts of the genotypic diversity in the target population will be under-represented in the training set, whereas other parts will be over-represented. Therefore, we propose a training set construction method that uniformly samples the genetic space spanned by the target population of genotypes, thereby increasing predictive ability. To evaluate our method, we constructed training sets alongside with the identification of corresponding genomic prediction models for four genotype panels that differed in the amount of population structure they contained (maize Flint, maize Dent, wheat, and rice). Training sets were constructed using uniform sampling, stratified-uniform sampling, stratified sampling and random sampling. We compared these methods with a method that maximizes the generalized coefficient of determination (CD). Several training set sizes were considered. We investigated four genomic prediction models: multi-locus QTL models, GBLUP models, combinations of QTL and GBLUPs, and Reproducing Kernel Hilbert Space (RKHS) models. For the maize and wheat panels, construction of the training set under uniform sampling led to a larger predictive ability than under stratified and random sampling. The results of our methods were similar to those of the CD method. For the rice panel, all training set construction methods led to similar predictive ability, a reflection of the very strong population structure in this panel.

A Fully Automatic Player Detection Method Based on One-Class SVM

Player detection is an important part in sports video analysis. Over the past few years, several learning based detection methods using various supervised two-class techniques have been presented. Although satisfactory results can be obtained, a lot of manual labor is needed to construct the training set. To overcome this drawback, this letter proposes a player detection method based on one-class SVM (OCSVM) using automatically generated training data. The proposed method is evaluated using several video clips captured from World Cup 2010, and experimental results show that our approach achieves a high detection rate while keeping the training set construction's cost low.

Label-specific training set construction from web resource for image annotation

Recently many research efforts have been devoted to image annotation by leveraging on the associated tags/keywords of web images as training labels. A key issue to resolve is the relatively low accuracy of the tags. In this paper, we propose a novel semi-automatic framework to construct a more accurate and effective training set from these web media resources for each label that we want to learn. Locality sensitive Hashing (LSH) is applied to find the most possible region candidates of a given label efficiently. We further conduct simple human interactions to approve whether the clusters of region candidates are relevant to the given label. Here Hashing ensures the efficiency and the minimal human efforts guarantee the effectiveness of the proposed framework. Experiments conducted on a real-world dataset demonstrate that the constructed training set can result in higher accuracy for image annotation.

Methods of training set construction: Towards improving performance for automated mesozooplankton image classification systems

The correspondence between variation in the physico-chemical properties of the water column and the taxonomic composition of zooplankton communities represents an important indicator of long-term and broad-scale change in marine systems. Evaluating and relating compositional change to various forms of perturbation demand routine taxonomic identification methods that can be applied rapidly and accurately. Traditional identification by human experts is accurate but very time-consuming. The application of automated image classification systems for plankton communities has emerged as a potential resolution to this limitation. The objective of this study is to evaluate how specific aspects of training set construction for the ZooScan system influenced our ability to relate variation in zooplankton taxonomic composition to variation of hydrographic properties in the East China Sea. Specifically, we compared the relative utility of zooplankton classifiers trained with the following: (i) water mass-specific and global training sets; (ii) balanced versus imbalanced training sets. The classification performance (accuracy and precision) of water-mass specific classifiers tended to decline with environmental dissimilarity, suggesting water-mass specificity However, similar classification performance was also achieved by training our system with samples representing all hydrographic sub-regions (i.e. a global classifier). After examining category-specific accuracy, we found that equal performance arises because the accuracy was mainly determined by dominant taxa. This apparently high classification accuracy was at the expense of accurate classification of rare taxa. To explore the basis for such biased classification, we trained our global classifier with an equal amount of training data for each category (balanced training). We found that balanced training had higher accuracy at recognizing rare taxa but low accuracy at abundant taxa. The errors introduced in recognition still pose a major challenge for automatic classification systems. In order to fully automate analyses of zooplankton communities and relate variation in composition to hydrographic properties, the recognition power of the system requires further improvements.

Sites Inferred by Metabolic Background Assertion Labeling (SIMBAL): adapting the Partial Phylogenetic Profiling algorithm to scan sequences for signatures that predict protein function

BackgroundComparative genomics methods such as phylogenetic profiling can mine powerful inferences from inherently noisy biological data sets. We introduce Sites Inferred by Metabolic Background Assertion Labeling (SIMBAL), a method that applies the Partial Phylogenetic Profiling (PPP) approach locally within a protein sequence to discover short sequence signatures associated with functional sites. The approach is based on the basic scoring mechanism employed by PPP, namely the use of binomial distribution statistics to optimize sequence similarity cutoffs during searches of partitioned training sets.ResultsHere we illustrate and validate the ability of the SIMBAL method to find functionally relevant short sequence signatures by application to two well-characterized protein families. In the first example, we partitioned a family of ABC permeases using a metabolic background property (urea utilization). Thus, the TRUE set for this family comprised members whose genome of origin encoded a urea utilization system. By moving a sliding window across the sequence of a permease, and searching each subsequence in turn against the full set of partitioned proteins, the method found which local sequence signatures best correlated with the urea utilization trait. Mapping of SIMBAL "hot spots" onto crystal structures of homologous permeases reveals that the significant sites are gating determinants on the cytosolic face rather than, say, docking sites for the substrate-binding protein on the extracellular face. In the second example, we partitioned a protein methyltransferase family using gene proximity as a criterion. In this case, the TRUE set comprised those methyltransferases encoded near the gene for the substrate RF-1. SIMBAL identifies sequence regions that map onto the substrate-binding interface while ignoring regions involved in the methyltransferase reaction mechanism in general. Neither method for training set construction requires any prior experimental characterization.ConclusionsSIMBAL shows that, in functionally divergent protein families, selected short sequences often significantly outperform their full-length parent sequence for making functional predictions by sequence similarity, suggesting avenues for improved functional classifiers. When combined with structural data, SIMBAL affords the ability to localize and model functional sites.

Optimum Conditions for Construction of Training Set in Fringe Analysis with Multilayer Neural Network

A method of automatically processing a deformed fringe pattern in fringe projection profilometry to profile a three-dimensional (3D) object with a multilayer neural network has been investigated by computer simulation. The optimum projection conditions for the fringe pattern and the determination conditions for a training set for a learning process of the neural network were proposed to apply the multilayer neural network effectively. It is shown that the learning time in the neural network is reduced and that the performance of the profilometry is improved by the proposed optimization.