Classifier Training Research Articles

Багатоетапний метод глибинного навчання з попереднім самонавчанням для класифікаційного аналізу дефектів стічних труб

A machine learningsemi-supervised method was developed for the classification analysis of defects on the surface of the sewer pipe based on CCTV video inspection images. The aim of the research is the process of defect detection on the surface of sewage pipes. The subject of the research is a machine learning method for the classification analysis of sewage pipe defects on video inspection images under conditions of a limited and unbalanced set of labeled training data. A five-stage algorithm for classifier training is proposed. In the first stage, contrast training occurs using the instance-prototype contrast loss function, where the normalized Euclidean distance is used to measure the similarity of the encoded samples. The second step considers two variants of regularized loss functions – a triplet NCA function and a contrast-center loss function. The regularizing component in the second stage of training is used to penalize the rounding error of the output feature vector to a discrete form and ensures that the principle of information bottlenecking is implemented. The next step is to calculate the binary code of each class to implement error-correcting codes, but considering the structure of the classes and the relationships between their features. The resulting prototype vector of each class is used as a label of image for training using the cross-entropy loss function. The last stage of training conducts an optimization of the parameters of the decision rules using the information criterion to consider the variance of the class distribution in Hamming binary space. A micro-averaged metric F1, which is calculated on test data, is used to compare learning outcomes at different stages and within different approaches. The results obtained on the Sewer-ML open dataset confirm the suitability of the training method for practical use, with an F1 metric value of 0.977. The proposed method provides a 9 % increase in the value of the micro-averaged F1 metric compared to the results obtained using the traditional method.

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.

Development of Software for Control of Virtual Objects Using Event-Related Potential P300

The article considers the principle of operation of brain-computer interfaces (BCI) and a method for detecting the focus of a person's attention using event-related potential (P300). The review of the existing hardware and software solutions for the implementation of BCIs was performed including the identification of their advantages and disadvantages. The program that allows you to choose the desired stimulus from a variety of presented was developed.An electroencephalograph of the BiTronics Lab company on the Arduino platform was used to receive the signal. Signal filtering, classifier training and visualization are implemented in Python.

Realizing the Application of EEG Modeling in BCI Classification: Based on a Conditional GAN Converter.

Electroencephalogram (EEG) modeling in brain-computer interface (BCI) provides a theoretical foundation for its development. However, limited by the lack of guidelines in model parameter selection and the inability to obtain personal tissue information in practice, EEG modeling in BCI is mainly focused on the theoretical qualitative level which shows a gap between the theory and its application. Based on such problems, this work combined the surface EEG simulation with a converter based on the generative adversarial network (GAN), to establish the connection from simulated EEG to its application in BCI classification. For the scalp EEGs modeling, a mathematical model was built according to the physics of surface EEG, which consisted of the parallel 3-population neural mass model, the equivalent dipole, and the forward computation. For application, a converter based on the conditional GAN was designed, to transfer the simulated theoretical-only EEG to its practical version, in the lack of individual bio-information. To verify the feasibility, based on the latest microexpression-assisted BCI paradigm proposed by our group, the converted simulated EEGs were used in the training of BCI classifiers. The results indicated that, compared with training with insufficient real data, by adding the simulated EEGs, the overall performance showed a significant improvement (P = 0.04 < 0.05), and the test performance can be improved by 2.17% ± 4.23, in which the largest increase was up to 12.60% ± 1.81. Through this work, the link from theoretical EEG simulation to BCI classification has been initially established, providing an enhanced novel solution for the application of EEG modeling in BCI.

Temporomandibular Joint Osteoarthritis Diagnosis Using Privileged Learning of Protein Markers.

Diagnosis of temporomandibular joint (TMJ) Osteoarthritis (OA) before serious degradation of cartilage and subchondral bone occurs can help prevent chronic pain and disability. Clinical, radiomic, and protein markers collected from TMJ OA patients have been shown to be predictive of OA onset. Since protein data can often be unavailable for clinical diagnosis, we harnessed the learning using privileged information (LUPI) paradigm to make use of protein markers only during classifier training. Three different LUPI algorithms are compared with traditional machine learning models on a dataset extracted from 92 unique OA patients and controls. The best classifier performance of 0.80 AUC and 75.6 accuracy was obtained from the KRVFL+ model using privileged protein features. Results show that LUPI-based algorithms using privileged protein data can improve final diagnostic performance of TMJ OA classifiers without needing protein microarray data during classifier diagnosis.

Machine Learning-Based Automated Selection of Regions for Analysis on Bone Marrow Aspirate Smears

Abstract Manual microscopic examination of bone marrow aspirate (BMA) smears and counting of cell populations remains the standard of practice for accurate assessment of benign and neoplastic bone marrow disorders. While automated cell classification software using machine learning models has been developed and applied to BMAs, current systems nonetheless require manual identification of optimal regions within the slide that are rich in marrow hematopoietic cells. To address this issue, we have developed a machine learning-based platform for automated identification of optimal regions in whole-slide images of BMA smears. A training dataset was developed by manual annotation of 53 BMA slides across biopsy diagnoses including unremarkable trilineal hematopoiesis, acute leukemia, and plasma cell neoplasms, as well as across differences in total cellularity represented by a spectrum of marrow nucleated cell content and white blood cell counts. 10,537 regions among these 53 slides were manually annotated as either “optimal” (regions near aspirate particles with high proportions of marrow nucleated cells), “particle” (aspirate particles), or “hemodilute” (blood-rich regions with high proportions of red blood cells). Training of a neural network-based classifier on 10x magnification slides with region cropping and image augmentation resulted in a classifier with substantial accuracy on new testing-set BMA slides (one-vs-rest AUROC &amp;gt; 0.999 across 10 training/testing splits for all 3 region classes), with very few particle and hemodilute regions being classified as optimal (particle: 0.83%, hemodilute: 0.39%). Additionally, this classifier accurately classifies BMA regions on slides from hematological disorders not represented in the training data, including Burkitt lymphoma (AUROC &amp;gt; 0.999 across region classes), chronic myeloid leukemia (AUROC &amp;gt; 0.999 across region classes), and diffuse large B-cell lymphoma (AUROC = 1 across region classes), demonstrating the broad applicability of our approach. To assess the performance of our classifier on whole-slide images, tiles from 10x magnification slides were manually annotated by three participants with notable concordance (Krippendorff’s alpha = 0.424); substantial agreement was found between manual annotations and model predictions within whole-slide images (optimal AUROC = 0.958, particle AUROC = 1.0, hemodilute AUROC = 0.947). Based on these promising results, this machine learning-based region classification model is being connected to a previously-developed bone marrow cell classifier to fully automate differential cell counting in whole-slide images. The development of this novel automated pipeline has potential to streamline the diagnostic process for hematological disorders while enhancing accuracy and replicability, as well as decreasing diagnostic turnaround time for improving patient care.

Rainfall observation using surveillance audio

Environmental audio recordings are a rich and underexploited source of rainfall events. Widespread surveillance cameras continuously record rainfall information, which provides a basis for the possibility of rainfall monitoring. In this study, using surveillance audio as input, an automatic rainfall level classification system was built. Through rainfall level definitions, the rainfall observation task can be transformed into an audio classification task. Three 2-D baseline convolutional neural networks (CNNs) were proposed as the classifiers. In view of the classifier training and testing, a new dataset named Rainfall Audio_XZ (RA_XZ) was generated based on the surveillance audio data. Because of the significant impact of the input features on the classification performance, 16 kinds of feature aggregation schemes were constructed and investigated to find the best representation of rainfall. The experimental results demonstrate that the proposed CNN (7-stack CNN) achieves 81.67% accuracy in rainfall level classification in the RA_XZ dataset and 93.38% accuracy in environmental sound classification in the UrbanSound8k dataset, which is optimal performance compared to that of some existing relevant algorithms. Furthermore, audio aggregation strategies that facilitate the representation and classification of rainfall events are investigated, which has important implications for audio and speech classification systems. In summary, our study provides an effective supplement to traditional rainfall monitoring technologies.

On the importance of representative datasets in ECG-based artificial intelligence

Abstract Background/Introduction ECG-based artificial intelligence (AI) is an emerging field in digital cardiology. Training on diseased records vs. healthy controls is common practice. We aimed to evaluate if such an approach can lead to unwanted behaviour in real-world settings and thus unnecessarily reduce diagnostic precision of the developed AI model. Purpose Several studies have shown that deep neural networks are able to exceed performance of medical experts. However, when these models are applied to different cohorts, results vary strongly. We hypothesise that this is because the datasets used for training were not representative for the target population. Methods Based on the public ECG database PTB-XL we sampled three distinct subsets of n=150 records representing ECG groups labelled for diagnoses 'old myocardial infarction' (M), 'normal ECG' (N), or 'other cardiac abnormality' (O). These groups were combined to three datasets ([M, N] (n=300), [M, O] (n=300), [M, N, O] (n=450)), representing different approaches to data sampling. On each dataset, we trained a separate but equally structured deep neural network using 100-fold bootstrapping. The diagnostic performance of each model was validated on unseen data from all datasets with sensitivity, specificity and area under the receiver operator characteristic curve. Results Evaluation of the three differently trained models shows best diagnostic performance on the M vs. N records and worst on the M vs. O records. However, in the out-of-dataset setting, the best-performing model (trained on [M, N]) shows weaker performance on the [M, N, O] and [M, O] datasets. Sensitivity for the same model remained equal, as identical M records were used throughout corresponding bootstrapping folds. Detailed results are presented in Table 1. Conclusions Our results suggest that the model trained on a dataset including only diseased records vs. healthy controls [M, N] learned to recognise healthy (N) instead of diseased (M) records, which explains why it performed poorly on datasets including records showing other cardiac abnormalities (O). Such behaviour is a common problem in AI and requires special attention in dataset sampling. For small cohorts, it is tempting to increase the amount of training data by using healthy controls. However, we have shown that this can be a poor option, since classifiers can more easily rely on features that are not actually related to the target disease. Training and validation of classifiers should therefore be performed on representative datasets that are as close as possible to the target population. Funding Acknowledgement Type of funding sources: Public Institution(s). Main funding source(s): Forschungscampus Mittelhessen, Flexi Funds Table 1

A MODEL AND TRAINING METHOD FOR CONTEXT CLASSIFICATION IN CCTV SEWER INSPECTION VIDEO FRAMES

Context. A model and training method for observational context classification in CCTV sewer inspection vide frames was developed and researched. The object of research is the process of detection of temporal-spatial context during CCTV sewer inspections. The subjects of the research are machine learning model and training method for classification analysis of CCTV video sequences under the limited and imbalanced training dataset constraint.&#x0D; Objective. Stated research goal is to develop an efficient context classifier model and training algorithm for CCTV sewer inspection video frames under the constraint of the limited and imbalanced labeled training set.&#x0D; Methods. The four-stage training algorithm of the classifier is proposed. The first stage involves training with soft triplet loss and regularisation component which penalises the network’s binary output code rounding error. The next stage is needed to determine the binary code for each class according to the principles of error-correcting output codes with accounting for intra- and interclass relationship. The resulting reference vector for each class is then used as a sample label for the future training with Joint Binary Cross Entropy Loss. The last machine learning stage is related to decision rule parameter optimization according to the information criteria to determine the boundaries of deviation of binary representation of observations for each class from the corresponding reference vector. A 2D convolutional frame feature extractor combined with the temporal network for inter-frame dependency analysis is considered. Variants with 1D Dilated Regular Convolutional Network, 1D Dilated Causal Convolutional Network, LSTM Network, GRU Network are considered. Model efficiency comparison is made on the basis of micro averaged F1 score calculated on the test dataset.&#x0D; Results. Results obtained on the dataset provided by Ace Pipe Cleaning, Inc confirm the suitability of the model and method for practical use, the resulting accuracy equals 92%. Comparison of the training outcome with the proposed method against the conventional methods indicated a 4% advantage in micro averaged F1 score. Further analysis of the confusion matrix had shown that the most significant increase in accuracy in comparison with the conventional methods is achieved for complex classes which combine both camera orientation and the sewer pipe construction features.&#x0D; Conclusions. The scientific novelty of the work lies in the new models and methods of classification analysis of the temporalspatial context when automating CCTV sewer inspections under imbalanced and limited training dataset conditions. Training results obtained with the proposed method were compared with the results obtained with the conventional method. The proposed method showed 4% advantage in micro averaged F1 score.&#x0D; It had been empirically proven that the use of the regular convolutional temporal network architecture is the most efficient in utilizing inter-frame dependencies. Resulting accuracy is suitable for practical use, as the additional error correction can be made by using the odometer data.

Stability for the training of deep neural networks and other classifiers

We examine the stability of loss-minimizing training processes that are used for deep neural networks (DNN) and other classifiers. While a classifier is optimized during training through a so-called loss function, the performance of classifiers is usually evaluated by some measure of accuracy, such as the overall accuracy which quantifies the proportion of objects that are well classified. This leads to the guiding question of stability: does decreasing loss through training always result in increased accuracy? We formalize the notion of stability, and provide examples of instability. Our main result consists of two novel conditions on the classifier which, if either is satisfied, ensure stability of training, that is we derive tight bounds on accuracy as loss decreases. We also derive a sufficient condition for stability on the training set alone, identifying flat portions of the data manifold as potential sources of instability. The latter condition is explicitly verifiable on the training dataset. Our results do not depend on the algorithm used for training, as long as loss decreases with training.

Classification of Tumor of MRI Brain Image Using Hybrid Feature Extraction Method and Support Vector Machine Classifier

Brain tumors Analysis is problematic somewhat due to varied size, shape, location of tumor and the appearance and presence of brain tumor. Clinicians and radiologist have difficulty in identifying the tumor type. An efficient hybrid feature extraction method to classify the type of tumor accurately as meningioma, gliomas and pituitary tumor using SVM (support vector machine) classifier is proposed. The modified Non-Local Means (NLM) filter may be effectively used to get the pure image. The NLM filter is compared with common filters like median and wiener. From the denoised image the classification is done by training SVM using the texture features from the hybrid and efficient feature extraction technique.The accuracy of the classification is calculated and the SVM classifier training individual type of texture features and also with combined texture features and the performance is analyzed.

A causality based feature selection approach for data-driven dynamic security assessment

The integration of renewable energy sources increases the operational uncertainty of electric power systems and can lead to more frequent dynamic phenomena. The use of classifiers from machine learning is promising to include dynamics in the security assessment of the power system. The training of these classifiers is typically performed offline on synthetically generated operating conditions (OCs) that are similar to real-time operation. However, the uncertainty in the generated OCs and the classifier’s inaccuracy is larger the longer the time between offline and real-time operation. Moving the classifier training closer to real-time operation is an important step forward to reduce inaccurate predictions and improve reliability. In this paper, a novel causality-based feature selection approach for an online dynamic security assessment (DSA) framework is proposed. The key novelty is to use the system’s physics to learn the causal structure between the features and then select the features based on this causal structure. The proposed approach results in faster computations, is more robust and more interpretable. Moreover, classifiers can be trained closer to real-time operation which enhances the predictive performance. Through a case study using transient stability on the IEEE 68-bus system, the proposed method reduces computational time by 75% in comparison to state of the art feature selection techniques. The proposed workflow showed superior performance in accuracy and robustness against uncertainty compared to conventional machine learning approaches for DSA. The computational benefit was also projected to a dataset of the French transmission system where the approach has the potential to achieve computational savings of up-to two orders of magnitudes.

A Novel Luminance-Based Algorithm for Classification of Semi-Dark Images

Image classification of a visual scene based on visibility is significant due to the rise in readily available automated solutions. Currently, there are only two known spectrums of image visibility i.e., dark, and bright. However, normal environments include semi-dark scenarios. Hence, visual extremes that will lead to the accurate extraction of image features should be duly discarded. Fundamentally speaking there are two broad methods to perform visual scene-based image classification, i.e., machine learning (ML) methods and computer vision methods. In ML, the issues of insufficient data, sophisticated hardware and inadequate image classifier training time remain significant problems to be handled. These techniques fail to classify the visual scene-based images with high accuracy. The other alternative is computer vision (CV) methods, which also have major issues. CV methods do provide some basic procedures which may assist in such classification but, to the best of our knowledge, no CV algorithm exists to perform such classification, i.e., these do not account for semi-dark images in the first place. Moreover, these methods do not provide a well-defined protocol to calculate images’ content visibility and thereby classify images. One of the key algorithms for calculation of images’ content visibility is backed by the HSL (hue, saturation, lightness) color model. The HSL color model allows the visibility calculation of a scene by calculating the lightness/luminance of a single pixel. Recognizing the high potential of the HSL color model, we propose a novel framework relying on the simple approach of the statistical manipulation of an entire image’s pixel intensities, represented by HSL color model. The proposed algorithm, namely, Relative Perceived Luminance Classification (RPLC) uses the HSL (hue, saturation, lightness) color model to correctly identify the luminosity values of the entire image. Our findings prove that the proposed method yields high classification accuracy (over 78%) with a small error rate. We show that the computational complexity of RPLC is much less than that of the state-of-the-art ML algorithms.

A low-cost and efficient electronic nose system for quantification of multiple indoor air contaminants utilizing HC and PLSR

A novel quantification technique is presented for electronic nose (E-nose), which is based on a double-step strategy combined with hierarchical classifier (HC) and partial least squares regression (PLSR). With the tree structure of HC, the complexity of classifier training process can be reduced in the case of unbalanced samples. For each level of the class hierarchy, the extreme learning machine-based artificial neural network (ELM-ANN) is applied for classification. In order to improve the classification performance of ELM-ANN, the multiple time-domain features are selected as training inputs, and a novel optimization method of the number of hidden layer neurons is given. To validate the effectiveness of the presented quantification technique, an E-nose system is designed to quantify the gases including six toxic gases (hydrogen sulfide, carbon monoxide, ammonia, toluene, formaldehyde, acetone) and three kinds of binary gas mixtures. This presented hierarchical classifier has demonstrated outstanding performance for the identification of target gases, such as the macro-averaged precision for unlabeled data is improved from 80% to 92% compared with non-hierarchical classifiers. Furthermore, an excellent performance of concentration estimation is obtained utilizing PLSR, where average values of the coefficient of determination for training and test samples are equal to 0.957 and 0.927, respectively. Overall, our work demonstrates that the proposed approach is applicable in E-nose-based odor quantification.

Correction to: Playing to distraction: towards a robust training of CNN classifiers through visual explanation techniques

Correction to: Playing to distraction: towards a robust training of CNN classifiers through visual explanation techniques

CellProfiler Analyst 3.0: accessible data exploration and machine learning for image analysis.

Image-based experiments can yield many thousands of individual measurements describing each object of interest, such as cells in microscopy screens. CellProfiler Analyst is a free, open-source software package designed for the exploration of quantitative image-derived data and the training of machine learning classifiers with an intuitive user interface. We have now released CellProfiler Analyst 3.0, which in addition to enhanced performance adds support for neural network classifiers, identifying rare object subsets, and direct transfer of objects of interest from visualization tools into the Classifier tool for use as training data. This release also increases interoperability with the recently released CellProfiler 4, making it easier for users to detect and measure particular classes of objects in their analyses. CellProfiler Analyst binaries for Windows and MacOS are freely available for download at https://cellprofileranalyst.org/. Source code is implemented in Python 3 and is available at https://github.com/CellProfiler/CellProfiler-Analyst/. A sample dataset is available at https://cellprofileranalyst.org/examples, based on images freely available from the Broad Bioimage Benchmark Collection.

Quantum federated learning through blind quantum computing

Private distributed learning studies the problem of how multiple distributed entities collaboratively train a shared deep network with their private data unrevealed. With the security provided by the protocols of blind quantum computation, the cooperation between quantum physics and machine learning may lead to unparalleled prospect for solving private distributed learning tasks. In this paper, we introduce a quantum protocol for distributed learning that is able to utilize the computational power of the remote quantum servers while keeping the private data safe. For concreteness, we first introduce a protocol for private single-party delegated training of variational quantum classifiers based on blind quantum computing and then extend this protocol to multiparty private distributed learning incorporated with differential privacy. We carry out extensive numerical simulations with different real-life datasets and encoding strategies to benchmark the effectiveness of our protocol. We find that our protocol is robust to experimental imperfections and is secure under the gradient attack after the incorporation of differential privacy. Our results show the potential for handling computationally expensive distributed learning tasks with privacy guarantees, thus providing a valuable guide for exploring quantum advantages from the security perspective in the field of machine learning with real-life applications.

Joint discriminative and shared dictionary learning with dictionary extension strategy for bearing fault classification

Sparse representation is one of the effective approaches for bearing fault diagnosis. Conventional sparse representation only focuses on fault feature extraction from bearing vibration signals while requiring experts to empirically make a diagnosis. This paper investigates bearing intelligent fault classification based on sparse representation based classification (SRC) method. A novel SRC algorithm, named joint discriminative and shared dictionary learning (JDSDL), is proposed for simultaneously characterize the class-specific and common features in vibration signals of bearings with different faults. The JDSDL seamlessly integrates discriminative dictionary learning, shared dictionary learning, and classifier training into a joint model to automatically recognize bearing faults. Moreover, a dictionary extension strategy is embedded within the proposed JDSDL model to make learned discriminative dictionary shift-invariant to improve its classification performance. The main advantages of the proposed method are that the JDSDL not only works well even if the training set is small, but also is robust to corruption of environment noises. Experimental results on three bearing datasets and comparison with competing methods are presented to demonstrate the effectiveness and superiority of the proposed JDSDL in bearing intelligent fault classification.

Brain Tumor Detection using a combination of Bayesian optimization based SVM classifier and fine-tuned based deep features

Brain tumor, one of the most common types of cancer, is a fatal disease. Therefore, accurate diagnosis of this disease and determining the type of tumor are of great importance in terms of early treatment. In this context, research, and interest in the development of automatic systems for the problems experienced in brain tumor classification, based on deep learning, have increased recently. In this study, a unique framework is proposed, which is based on Bayesian optimization-based Support Vector Machine (SVM) classifier and Convolutional Neural Network (CNN) based deep features ensemble, for the classification of brain tumors. In this model, brain MRI images are first improved. Second, the deep features are extracted using pre-trained CNN-based deep architectures and then combined. Later, effective, and distinctive features are selected from these deep features with the MrMr algorithm. Finally, these features are used in the training of the SVM classifier based on the Bayesian optimization algorithm. A dataset named Figshare, containing brain tumor images such as meningioma, glioma, and pituitary, is used to test the proposed system. In the experimental studies, the accuracy score of the model proposed was observed to be more successful than that of the other studies.

AI 딥러닝을 이용한 싸움(다툼) 행동 인식 방법

In this study, we developed an AI deep learning-based fighting behavior recognition method for a video surveillance system and proved its effectiveness through various experiments. The proposed method consists of a two-step fighting behavior recognition framework. First, continuous video frames of the target surveillance video are transmitted to the Inflated 3D ConvNet (I3D) network, which shows a good behavior-recognition performance, to extract the spatiotemporal features. These extracted 3D features are then used as the inputs in the next step, where a fight situation is detected using a classification model consisting of a fully connected layer. To use the proposed aggressive behavior detection framework effectively, first, it is necessary to train the fight detection model. However, it is not possible to collect sufficient fighting videos in various outdoor environments. To overcome this limitation, we generated a large amount of learning data through data augmentation. Therefore, instead of directly learning from the training videos transmitted to the I3D network, the classifier trains itself to recognize various fighting actions using the Kinetics video dataset. That is, the action features are extracted from the transmitted consecutive frames using the pretrained I3D network and subsequently used to train the fully connected layer classification model. In addition, we proposed a learning method that includes recognizing ambiguous conflict boundaries using multiple instance learning to mitigate the ambiguous starting and ending of the contention videos. The effectiveness of the proposed method was verified through several experiments by drawing comparisons between the present results and those of the previously reported studies.