Search results for: deep learning model

Authors: Jacob Green, Cecilia Cabrera, Maximilian Jakobs, Andrea Dimitracopoulos, Mark van der Wilk, Ryan Greenhalgh

Abstract:

Predicting drug properties is key in drug discovery to enable de-risking of assets before expensive clinical trials and to find highly active compounds faster. Interest from the machine learning community has led to the release of a variety of benchmark datasets and proposed methods. However, it remains unclear for practitioners which method or approach is most suitable, as different papers benchmark on different datasets and methods, leading to varying conclusions that are not easily compared. Our large-scale empirical study links together numerous earlier works on different datasets and methods, thus offering a comprehensive overview of the existing property classes, datasets, and their interactions with different methods. We emphasise the importance of uncertainty quantification and the time and, therefore, cost of applying these methods in the drug development decision-making cycle. To the best of the author's knowledge, it has been observed that the optimal approach varies depending on the dataset and that engineered features with classical machine learning methods often outperform deep learning. Specifically, QSAR datasets are typically best analysed with classical methods such as Gaussian Processes, while ADMET datasets are sometimes better described by Trees or deep learning methods such as Graph Neural Networks or language models. Our work highlights that practitioners do not yet have a straightforward, black-box procedure to rely on and sets a precedent for creating practitioner-relevant benchmarks. Deep learning approaches must be proven on these benchmarks to become the practical method of choice in drug property prediction.

Keywords: activity (QSAR), ADMET, classical methods, drug property prediction, empirical study, machine learning

Procedia PDF Downloads 81

Authors: Airy Huang, Fuji Foo, Aries Prasetya Wibowo

Abstract:

In recent years, the focus of Natural Language Processing (NLP) development has been gradually shifting from the semantics-based approach to deep learning one, which performs faster with fewer resources. Although it performs well in many applications, the deep learning approach, due to the lack of semantics understanding, has difficulties in noticing and expressing a novel business case with a pre-defined scope. In order to meet the requirements of specific robotic services, deep learning approach is very labor-intensive and time consuming. It is very difficult to improve the capabilities of conversational AI in a short time, and it is even more difficult to self-learn from experiences to deliver the same service in a better way. In this paper, we present an adaptive conversational AI algorithm that combines both semantic knowledge and deep learning to address this issue by learning new business cases through conversations. After self-learning from experience, the robot adapts to the business cases originally out of scope. The idea is to build new or extended robotic services in a systematic and fast-training manner with self-configured programs and constructed dialog flows. For every cycle in which a chat bot (conversational AI) delivers a given set of business cases, it is trapped to self-measure its performance and rethink every unknown dialog flows to improve the service by retraining with those new business cases. If the training process reaches a bottleneck and incurs some difficulties, human personnel will be informed of further instructions. He or she may retrain the chat bot with newly configured programs, or new dialog flows for new services. One approach employs semantics analysis to learn the dialogues for new business cases and then establish the necessary ontology for the new service. With the newly learned programs, it completes the understanding of the reaction behavior and finally uses dialog flows to connect all the understanding results and programs, achieving the goal of self-learning process. We have developed a chat bot service mounted on a kiosk, with a camera for facial recognition and a directional microphone array for voice capture. The chat bot serves as a concierge with polite conversation for visitors. As a proof of concept. We have demonstrated to complete 90% of reception services with limited self-learning capability.

Keywords: conversational AI, chatbot, dialog management, semantic analysis

Procedia PDF Downloads 136

Authors: Ola Hall, Ibrahim Wahab, Thorsteinn Rognvaldsson, Mattias Ohlsson

Abstract:

The subfield of poverty and welfare estimation that applies machine learning tools and methods on satellite imagery is a nascent but rapidly growing one. This is in part driven by the sustainable development goal, whose overarching principle is that no region is left behind. Among other things, this requires that welfare levels can be accurately and rapidly estimated at different spatial scales and resolutions. Conventional tools of household surveys and interviews do not suffice in this regard. While they are useful for gaining a longitudinal understanding of the welfare levels of populations, they do not offer adequate spatial coverage for the accuracy that is needed, nor are their implementation sufficiently swift to gain an accurate insight into people and places. It is this void that satellite imagery fills. Previously, this was near-impossible to implement due to the sheer volume of data that needed processing. Recent advances in machine learning, especially the deep learning subtype, such as deep neural networks, have made this a rapidly growing area of scholarship. Despite their unprecedented levels of performance, such models lack transparency and explainability and thus have seen limited downstream applications as humans generally are apprehensive of techniques that are not inherently interpretable and trustworthy. While several studies have demonstrated the superhuman performance of AI models, none has directly compared the performance of such models and human readers in the domain of poverty studies. In the present study, we directly compare the performance of human readers and a DL model using different resolutions of satellite imagery to estimate the welfare levels of demographic and health survey clusters in Tanzania, using the wealth quintile ratings from the same survey as the ground truth data. The cluster-level imagery covers all 608 cluster locations, of which 428 were classified as rural. The imagery for the human readers was sourced from the Google Maps Platform at an ultra-high resolution of 0.6m per pixel at zoom level 18, while that of the machine learning model was sourced from the comparatively lower resolution Sentinel-2 10m per pixel data for the same cluster locations. Rank correlation coefficients of between 0.31 and 0.32 achieved by the human readers were much lower when compared to those attained by the machine learning model – 0.69-0.79. This superhuman performance by the model is even more significant given that it was trained on the relatively lower 10-meter resolution satellite data while the human readers estimated welfare levels from the higher 0.6m spatial resolution data from which key markers of poverty and slums – roofing and road quality – are discernible. It is important to note, however, that the human readers did not receive any training before ratings, and had this been done, their performance might have improved. The stellar performance of the model also comes with the inevitable shortfall relating to limited transparency and explainability. The findings have significant implications for attaining the objective of the current frontier of deep learning models in this domain of scholarship – eXplainable Artificial Intelligence through a collaborative rather than a comparative framework.

Keywords: poverty prediction, satellite imagery, human readers, machine learning, Tanzania

Procedia PDF Downloads 104

Authors: Paulo Sérgio Ribeiro de Araújo Bogas

Abstract:

Although several studies have assumed (at least implicitly) that learners' approaches to learning develop into deeper approaches to higher education, there appears to be no clear theoretical basis for this assumption and no empirical evidence. As a scientific contribution to this discussion, a pedagogical intervention of a quasi-experimental nature was developed, with a mixed methodology, evaluating the intervention within a single curricular unit of Marketing, using cases based on real challenges of brands, business simulation, and customer projects. Primary and secondary experiences were incorporated in the intervention: the primary experiences are the experiential activities themselves; the secondary experiences result from the primary experience, such as reflection and discussion in work teams. A diversified learning relationship was encouraged through the various connections between the different members of the learning community. The present study concludes that in the same context, the student's responses can be described as students who reinforce the initial deep approach, students who maintain the initial deep approach level, and others who change from an emphasis on the deep approach to one closer to superficial. This typology did not always confirm studies reported in the literature, namely, whether the initial level of deep processing would influence the superficial and the opposite. The result of this investigation points to the inclusion of pedagogical and didactic activities that integrate different motivations and initial strategies, leading to the possible adoption of deep approaches to learning since it revealed statistically significant differences in the difference in the scores of the deep/superficial approach and the experiential level. In the case of real challenges, the categories of “attribution of meaning and meaning of studied” and the possibility of “contact with an aspirational context” for their future professional stand out. In this category, the dimensions of autonomy that will be required of them were also revealed when comparing the classroom context of real cases and the future professional context and the impact they may have on the world. Regarding the simulated practice, two categories of response stand out: on the one hand, the motivation associated with the possibility of measuring the results of the decisions taken, an awareness of oneself, and, on the other hand, the additional effort that this practice required for some of the students.

Keywords: experiential learning, higher education, mixed methods, reflective learning, marketing

Procedia PDF Downloads 83

Authors: Shenlun Chen, Leonard Wee

Abstract:

Tumor grading is an essential reference for colorectal cancer (CRC) staging and survival prognostication. The widely used World Health Organization (WHO) grading system defines histological grade of CRC adenocarcinoma based on the density of glandular formation on whole slide images (WSI). Tumors are classified as well-, moderately-, poorly- or un-differentiated depending on the percentage of the tumor that is gland forming; >95%, 50-95%, 5-50% and <5%, respectively. However, manually grading WSIs is a time-consuming process and can cause observer error due to subjective judgment and unnoticed regions. Furthermore, pathologists’ grading is usually coarse while a finer and continuous differentiation grade may help to stratifying CRC patients better. In this study, a deep learning based automatic differentiation grading algorithm was developed and evaluated by survival analysis. Firstly, a gland segmentation model was developed for segmenting gland structures. Gland regions of WSIs were delineated and used for differentiation annotating. Tumor regions were annotated by experienced pathologists into high-, medium-, low-differentiation and normal tissue, which correspond to tumor with clear-, unclear-, no-gland structure and non-tumor, respectively. Then a differentiation prediction model was developed on these human annotations. Finally, all enrolled WSIs were processed by gland segmentation model and differentiation prediction model. The differentiation grade can be calculated by deep learning models’ prediction of tumor regions and tumor differentiation status according to WHO’s defines. If multiple WSIs were possessed by a patient, the highest differentiation grade was chosen. Additionally, the differentiation grade was normalized into scale between 0 to 1. The Cancer Genome Atlas, project COAD (TCGA-COAD) project was enrolled into this study. For the gland segmentation model, receiver operating characteristic (ROC) reached 0.981 and accuracy reached 0.932 in validation set. For the differentiation prediction model, ROC reached 0.983, 0.963, 0.963, 0.981 and accuracy reached 0.880, 0.923, 0.668, 0.881 for groups of low-, medium-, high-differentiation and normal tissue in validation set. Four hundred and one patients were selected after removing WSIs without gland regions and patients without follow up data. The concordance index reached to 0.609. Optimized cut off point of 51% was found by “Maxstat” method which was almost the same as WHO system’s cut off point of 50%. Both WHO system’s cut off point and optimized cut off point performed impressively in Kaplan-Meier curves and both p value of logrank test were below 0.005. In this study, gland structure of WSIs and differentiation status of tumor regions were proven to be predictable through deep leaning method. A finer and continuous differentiation grade can also be automatically calculated through above models. The differentiation grade was proven to stratify CAC patients well in survival analysis, whose optimized cut off point was almost the same as WHO tumor grading system. The tool of automatically calculating differentiation grade may show potential in field of therapy decision making and personalized treatment.

Keywords: colorectal cancer, differentiation, survival analysis, tumor grading

Procedia PDF Downloads 134

Authors: Huda Al Shuaily, Karen Renaud

Abstract:

Databases comprise the foundation of most software systems. System developers inevitably write code to query these databases. The de facto language for querying is SQL and this, consequently, is the default language taught by higher education institutions. There is evidence that learners find it hard to master SQL, harder than mastering other programming languages such as Java. Educators do not agree about explanations for this seeming anomaly. Further investigation may well reveal the reasons. In this paper, we report on our investigations into how novices learn SQL, the actual problems they experience when writing SQL, as well as the differences between expert and novice SQL query writers. We conclude by presenting a model of SQL learning that should inform the instructional material design process better to support the SQL learning process.

Keywords: pattern, SQL, learning, model

Procedia PDF Downloads 254

Authors: Hongmei Wang, Ziyun Xiang, Ying liu, Li Yu, Dongsheng Yue

Abstract:

Intro: The COVID-19 (Corona Virus Disease 2019) has greatly changed the global economic, political and financial ecology. The mutation of the coronavirus in the UK in December 2020 has brought new panic to the world. Deep learning was performed on Chest Computed tomography (CT) of COVID-19 and Influenza and describes their characteristics. The predominant features of COVID-19 pneumonia was ground-glass opacification, followed by consolidation. Lesion density: most lesions appear as ground-glass shadows, and some lesions coexist with solid lesions. Lesion distribution: the focus is mainly on the dorsal side of the periphery of the lung, with the lower lobe of the lungs as the focus, and it is often close to the pleura. Other features it has are grid-like shadows in ground glass lesions, thickening signs of diseased vessels, air bronchi signs and halo signs. The severe disease involves whole bilateral lungs, showing white lung signs, air bronchograms can be seen, and there can be a small amount of pleural effusion in the bilateral chest cavity. At the same time, this year's flu season could be near its peak after surging throughout the United States for months. Chest CT for Influenza infection is characterized by focal ground glass shadows in the lungs, with or without patchy consolidation, and bronchiole air bronchograms are visible in the concentration. There are patchy ground-glass shadows, consolidation, air bronchus signs, mosaic lung perfusion, etc. The lesions are mostly fused, which is prominent near the hilar and two lungs. Grid-like shadows and small patchy ground-glass shadows are visible. Deep neural networks have great potential in image analysis and diagnosis that traditional machine learning algorithms do not. Method: Aiming at the two major infectious diseases COVID-19 and influenza, which are currently circulating in the world, the chest CT of patients with two infectious diseases is classified and diagnosed using deep learning algorithms. The residual network is proposed to solve the problem of network degradation when there are too many hidden layers in a deep neural network (DNN). The proposed deep residual system (ResNet) is a milestone in the history of the Convolutional neural network (CNN) images, which solves the problem of difficult training of deep CNN models. Many visual tasks can get excellent results through fine-tuning ResNet. The pre-trained convolutional neural network ResNet is introduced as a feature extractor, eliminating the need to design complex models and time-consuming training. Fastai is based on Pytorch, packaging best practices for in-depth learning strategies, and finding the best way to handle diagnoses issues. Based on the one-cycle approach of the Fastai algorithm, the classification diagnosis of lung CT for two infectious diseases is realized, and a higher recognition rate is obtained. Results: A deep learning model was developed to efficiently identify the differences between COVID-19 and influenza using chest CT.

Keywords: COVID-19, Fastai, influenza, transfer network

Procedia PDF Downloads 142

Authors: A. Degale Desta, Cheng Jian

Abstract:

Deep learning application in computer vision is rapidly advancing, giving it the ability to monitor the public and quickly identify potentially anomalous behaviour from crowd scenes. Therefore, the purpose of the current work is to improve the performance of safety of people in crowd events from panic behaviour through introducing the innovative idea of Aggregation of Ensembles (AOE), which makes use of the pre-trained ConvNets and a pool of classifiers to find anomalies in video data with packed scenes. According to the theory of algorithms that applied K-means, KNN, CNN, SVD, and Faster-CNN, YOLOv5 architectures learn different levels of semantic representation from crowd videos; the proposed approach leverages an ensemble of various fine-tuned convolutional neural networks (CNN), allowing for the extraction of enriched feature sets. In addition to the above algorithms, a long short-term memory neural network to forecast future feature values and a handmade feature that takes into consideration the peculiarities of the crowd to understand human behavior. On well-known datasets of panic situations, experiments are run to assess the effectiveness and precision of the suggested method. Results reveal that, compared to state-of-the-art methodologies, the system produces better and more promising results in terms of accuracy and processing speed.

Keywords: action recognition, computer vision, crowd detecting and tracking, deep learning

Procedia PDF Downloads 161

Authors: Radhika Ranjan Roy

Abstract:

Deployment of machine learning (ML)/deep learning (DL) algorithms for cyberattack detection in operational communications networks (wireless and/or wire-line) is being delayed because of low-performance parameters (e.g., recall, precision, and f₁-score). If datasets become imbalanced, which is the usual case for communications networks, the performance tends to become worse. Complexities in handling reducing dimensions of the feature sets for increasing performance are also a huge problem. Mahalanobis algorithms have been widely applied in scientific research because Mahalanobis distance metric learning is a successful framework. In this paper, we have investigated the Mahalanobis binary classifier algorithm for increasing cyberattack detection performance over communications networks as a proof of concept. We have also found that high-dimensional information in intermediate features that are not utilized as much for classification tasks in ML/DL algorithms are the main contributor to the state-of-the-art of improved performance of the Mahalanobis method, even for imbalanced and sparse datasets. With no feature reduction, MD offers uniform results for precision, recall, and f₁-score for unbalanced and sparse NSL-KDD datasets.

Keywords: Mahalanobis distance, machine learning, deep learning, NS-KDD, local intrinsic dimensionality, chi-square, positive semi-definite, area under the curve

Procedia PDF Downloads 78

Authors: Mst Shapna Akter, Hossain Shahriar

Abstract:

One of the most important challenges in the field of software code audit is the presence of vulnerabilities in software source code. Every year, more and more software flaws are found, either internally in proprietary code or revealed publicly. These flaws are highly likely exploited and lead to system compromise, data leakage, or denial of service. C and C++ open-source code are now available in order to create a largescale, machine-learning system for function-level vulnerability identification. We assembled a sizable dataset of millions of opensource functions that point to potential exploits. We developed an efficient and scalable vulnerability detection method based on deep neural network models that learn features extracted from the source codes. The source code is first converted into a minimal intermediate representation to remove the pointless components and shorten the dependency. Moreover, we keep the semantic and syntactic information using state-of-the-art word embedding algorithms such as glove and fastText. The embedded vectors are subsequently fed into deep learning networks such as LSTM, BilSTM, LSTM-Autoencoder, word2vec, BERT, and GPT-2 to classify the possible vulnerabilities. Furthermore, we proposed a neural network model which can overcome issues associated with traditional neural networks. Evaluation metrics such as f1 score, precision, recall, accuracy, and total execution time have been used to measure the performance. We made a comparative analysis between results derived from features containing a minimal text representation and semantic and syntactic information. We found that all of the deep learning models provide comparatively higher accuracy when we use semantic and syntactic information as the features but require higher execution time as the word embedding the algorithm puts on a bit of complexity to the overall system.

Keywords: cyber security, vulnerability detection, neural networks, feature extraction

Procedia PDF Downloads 89

Authors: Shayan Mohajer Hamidi

Abstract:

Federated learning has emerged as a promising approach for distributed model training while preserving data privacy. However, the challenges of communication overhead, limited network resources, and slow convergence hinder its widespread adoption. On the other hand, knowledge distillation has shown great potential in compressing large models into smaller ones without significant loss in performance. In this paper, we propose an innovative framework that combines federated learning and knowledge distillation to address these challenges and enhance the efficiency of distributed learning. Our approach, called Federated Knowledge Distillation (FKD), enables multiple clients in a federated learning setting to collaboratively distill knowledge from a teacher model. By leveraging the collaborative nature of federated learning, FKD aims to improve model compression while maintaining privacy. The proposed framework utilizes a coded teacher model that acts as a reference for distilling knowledge to the client models. To demonstrate the effectiveness of FKD, we conduct extensive experiments on various datasets and models. We compare FKD with baseline federated learning methods and standalone knowledge distillation techniques. The results show that FKD achieves superior model compression, faster convergence, and improved performance compared to traditional federated learning approaches. Furthermore, FKD effectively preserves privacy by ensuring that sensitive data remains on the client devices and only distilled knowledge is shared during the training process. In our experiments, we explore different knowledge transfer methods within the FKD framework, including Fine-Tuning (FT), FitNet, Correlation Congruence (CC), Similarity-Preserving (SP), and Relational Knowledge Distillation (RKD). We analyze the impact of these methods on model compression and convergence speed, shedding light on the trade-offs between size reduction and performance. Moreover, we address the challenges of communication efficiency and network resource utilization in federated learning by leveraging the knowledge distillation process. FKD reduces the amount of data transmitted across the network, minimizing communication overhead and improving resource utilization. This makes FKD particularly suitable for resource-constrained environments such as edge computing and IoT devices. The proposed FKD framework opens up new avenues for collaborative and privacy-preserving distributed learning. By combining the strengths of federated learning and knowledge distillation, it offers an efficient solution for model compression and convergence speed enhancement. Future research can explore further extensions and optimizations of FKD, as well as its applications in domains such as healthcare, finance, and smart cities, where privacy and distributed learning are of paramount importance.

Keywords: federated learning, knowledge distillation, knowledge transfer, deep learning

Procedia PDF Downloads 75

Authors: Milagros Jaén-Vargas, Javier García Martínez, Karla Miriam Reyes Leiva, María Fernanda Trujillo-Guerrero, Francisco Fernandes, Sérgio Barroso Gonçalves, Miguel Tavares Silva, Daniel Simões Lopes, José Javier Serrano Olmedo

Abstract:

Human Activity Recognition (HAR) is having a growing impact in creating new applications and is responsible for emerging new technologies. Also, the use of wearable sensors is an important key to exploring the human body's behavior when performing activities. Hence, the use of these dispositive is less invasive and the person is more comfortable. In this study, a database that includes three activities is used. The activities were acquired from inertial measurement unit sensors (IMU) and motion capture systems (MOCAP). The main objective is differentiating the performance from four Deep Learning (DL) models: Deep Neural Network (DNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) and hybrid model Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM), when considering acceleration, velocity and position and evaluate if integrating the IMU acceleration to obtain velocity and position represent an increment in performance when it works as input to the DL models. Moreover, compared with the same type of data provided by the MOCAP system. Despite the acceleration data is cleaned when integrating, results show a minimal increase in accuracy for the integrated signals.

Keywords: HAR, IMU, MOCAP, acceleration, velocity, position, feature maps

Procedia PDF Downloads 98

Authors: Kumar Divyendra

Abstract:

Underwater Sensor Networks (UWSNs) have wide applications in areas such as water quality monitoring, marine wildlife management etc. A typical UWSN system consists of a set of sensors deployed randomly underwater which communicate with each other using acoustic links. RF communication doesn't work underwater, and GPS too isn't available underwater. Additionally Automated Underwater Vehicles (AUVs) are deployed to collect data from some special nodes called Cluster Heads (CHs). These CHs aggregate data from their neighboring nodes and forward them to the AUVs using optical links when an AUV is in range. This helps reduce the number of hops covered by data packets and helps conserve energy. We consider the three-dimensional model of the UWSN. Nodes are initially deployed randomly underwater. They attach themselves to the surface using a rod and can only move upwards or downwards using a pump and bladder mechanism. We use graph theory concepts to maximize the coverage volume while every node maintaining connectivity with at least one surface node. We treat the surface nodes as landmarks and each node finds out its hop distance from every surface node. We treat these hop-distances as coordinates and use them for AUV navigation. An AUV intending to move closer to a node with given coordinates moves hop by hop through nodes that are closest to it in terms of these coordinates. In absence of GPS, multiple different approaches like Inertial Navigation System (INS), Doppler Velocity Log (DVL), computer vision-based navigation, etc., have been proposed. These systems have their own drawbacks. INS accumulates error with time, vision techniques require prior information about the environment. We propose a method that makes use of the earth's magnetic field values for navigation and combines it with other methods that simultaneously increase the coverage volume under the UWSN. The AUVs are fitted with magnetometers that measure the magnetic intensity (I), horizontal inclination (H), and Declination (D). The International Geomagnetic Reference Field (IGRF) is a mathematical model of the earth's magnetic field, which provides the field values for the geographical coordinateson earth. Researchers have developed an inverse deep learning model that takes the magnetic field values and predicts the location coordinates. We make use of this model within our work. We combine this with with the hop-by-hop movement described earlier so that the AUVs move in such a sequence that the deep learning predictor gets trained as quickly and precisely as possible We run simulations in MATLAB to prove the effectiveness of our model with respect to other methods described in the literature.

Keywords: clustering, deep learning, network backbone, parallel computing

Procedia PDF Downloads 98

Authors: Tong He, Long Chen, Irag Mantegh, Wen-Fang Xie

Abstract:

This paper proposes an image-based obstacle avoidance and tracking target identification strategy in GPS-degraded or GPS-denied environment for an Unmanned Aerial Vehicle (UAV). The traditional force algorithm for obstacle avoidance could produce local minima area, in which UAV cannot get away obstacle effectively. In order to eliminate it, an artificial potential approach based on harmonic potential is proposed to guide the UAV to avoid the obstacle by using the vision system. And image-based visual servoing scheme (IBVS) has been adopted to implement the proposed obstacle avoidance approach. In IBVS, the pixel accuracy is a key factor to realize the obstacle avoidance. In this paper, the deep reinforcement learning framework has been applied by reducing pixel errors through constant interaction between the environment and the agent. In addition, the combination of OpenTLD and Tensorflow based on neural network is used to identify the type of tracking target. Numerical simulation in Matlab and ROS GAZEBO show the satisfactory result in target identification and obstacle avoidance.

Keywords: image-based visual servoing, obstacle avoidance, tracking target identification, deep reinforcement learning, artificial potential approach, neural network

Procedia PDF Downloads 143

Authors: Junjie Zhong, Kai Hong, Lei Wang

Abstract:

Passage Retrieval (PR) plays an important role in many Natural Language Processing (NLP) tasks. Traditional efficient retrieval models relying on exact term-matching, such as TF-IDF or BM25, have nowadays been exceeded by pre-trained language models which match by semantics. Though they gain effectiveness, deep language models often require large memory as well as time cost. To tackle the trade-off between efficiency and effectiveness in PR, this paper proposes Graph Passage Retriever (GraphPR), a graph-based model inspired by the development of graph learning techniques. Different from existing works, GraphPR is end-to-end and integrates both term-matching information and semantics. GraphPR constructs a passage-level graph from BM25 retrieval results and trains a GCN-like model on the graph with graph-based objectives. Passages were regarded as nodes in the constructed graph and were embedded in dense vectors. PR can then be implemented using embeddings and a fast vector-similarity search. Experiments on a variety of real-world retrieval datasets show that the proposed model outperforms related models in several evaluation metrics (e.g., mean reciprocal rank, accuracy, F1-scores) while maintaining a relatively low query latency and memory usage.

Keywords: efficiency, effectiveness, graph learning, language model, passage retrieval, term-matching model

Procedia PDF Downloads 148

Authors: Luther Ollier, Sylvie Thiria, Anastase Charantonis, Carlos E. Mejia, Michel Crépon

Abstract:

Sea Surface Height Anomaly (SLA) is a signature of the sub-mesoscale dynamics of the upper ocean. Sea Surface Temperature (SST) is driven by these dynamics and can be used to improve the spatial interpolation of SLA fields. In this study, we focused on the temporal evolution of SLA fields. We explored the capacity of deep learning (DL) methods to predict short-term SLA fields using SST fields. We used simulated daily SLA and SST data from the Mercator Global Analysis and Forecasting System, with a resolution of (1/12)◦ in the North Atlantic Ocean (26.5-44.42◦N, -64.25–41.83◦E), covering the period from 1993 to 2019. Using a slightly modified image-to-image convolutional DL architecture, we demonstrated that SST is a relevant variable for controlling the SLA prediction. With a learning process inspired by the teaching-forcing method, we managed to improve the SLA forecast at five days by using the SST fields as additional information. We obtained predictions of a 12 cm (20 cm) error of SLA evolution for scales smaller than mesoscales and at time scales of 5 days (20 days), respectively. Moreover, the information provided by the SST allows us to limit the SLA error to 16 cm at 20 days when learning the trajectory.

Keywords: deep-learning, altimetry, sea surface temperature, forecast

Procedia PDF Downloads 90

Authors: Yaolin Tian, Shanxiong Chen, Fujia Zhao, Xiaoyu Lin, Hailing Xiong

Abstract:

Ancient books are significant culture inheritors and their background textures convey the potential history information. However, multi-style texture recovery of ancient books has received little attention. Restricted by insufficient ancient textures and complex handling process, the generation of ancient textures confronts with new challenges. For instance, training without sufficient data usually brings about overfitting or mode collapse, so some of the outputs are prone to be fake. Recently, image generation and style transfer based on deep learning are widely applied in computer vision. Breakthroughs within the field make it possible to conduct research upon multi-style texture recovery of ancient books. Under the circumstances, we proposed a network of layout analysis and image fusion system. Firstly, we trained models by using Deep Convolution Generative against Networks (DCGAN) to synthesize multi-style ancient textures; then, we analyzed layouts based on the Position Rearrangement (PR) algorithm that we proposed to adjust the layout structure of foreground content; at last, we realized our goal by fusing rearranged foreground texts and generated background. In experiments, diversified samples such as ancient Yi, Jurchen, Seal were selected as our training sets. Then, the performances of different fine-turning models were gradually improved by adjusting DCGAN model in parameters as well as structures. In order to evaluate the results scientifically, cross entropy loss function and Fréchet Inception Distance (FID) are selected to be our assessment criteria. Eventually, we got model M8 with lowest FID score. Compared with DCGAN model proposed by Radford at el., the FID score of M8 improved by 19.26%, enhancing the quality of the synthetic images profoundly.

Keywords: deep learning, image fusion, image generation, layout analysis

Procedia PDF Downloads 156

Authors: Edward W. Staley, Corban G. Rivera, Ashley J. Llorens

Abstract:

Advances in reinforcement learning (RL) have resulted in recent breakthroughs in the application of artificial intelligence (AI) across many different domains. An emerging landscape of development environments is making powerful RL techniques more accessible for a growing community of researchers. However, most existing frameworks do not directly address the problem of learning in complex operating environments, such as dense urban settings or defense-related scenarios, that incorporate distributed, heterogeneous teams of agents. To help enable AI research for this important class of applications, we introduce the AI Arena: a scalable framework with flexible abstractions for distributed multi-agent reinforcement learning. The AI Arena extends the OpenAI Gym interface to allow greater flexibility in learning control policies across multiple agents with heterogeneous learning strategies and localized views of the environment. To illustrate the utility of our framework, we present experimental results that demonstrate performance gains due to a distributed multi-agent learning approach over commonly-used RL techniques in several different learning environments.

Keywords: reinforcement learning, multi-agent, deep learning, artificial intelligence

Procedia PDF Downloads 157

Authors: Alexander Guzman Urbina, Atsushi Aoyama

Abstract:

The sustainability of traditional technologies employed in energy and chemical infrastructure brings a big challenge for our society. Making decisions related with safety of industrial infrastructure, the values of accidental risk are becoming relevant points for discussion. However, the challenge is the reliability of the models employed to get the risk data. Such models usually involve large number of variables and with large amounts of uncertainty. The most efficient techniques to overcome those problems are built using Artificial Intelligence (AI), and more specifically using hybrid systems such as Neuro-Fuzzy algorithms. Therefore, this paper aims to introduce a hybrid algorithm for risk assessment trained using near-miss accident data. As mentioned above the sustainability of traditional technologies related with energy and chemical infrastructure constitutes one of the major challenges that today’s societies and firms are facing. Besides that, the adaptation of those technologies to the effects of the climate change in sensible environments represents a critical concern for safety and risk management. Regarding this issue argue that social consequences of catastrophic risks are increasing rapidly, due mainly to the concentration of people and energy infrastructure in hazard-prone areas, aggravated by the lack of knowledge about the risks. Additional to the social consequences described above, and considering the industrial sector as critical infrastructure due to its large impact to the economy in case of a failure the relevance of industrial safety has become a critical issue for the current society. Then, regarding the safety concern, pipeline operators and regulators have been performing risk assessments in attempts to evaluate accurately probabilities of failure of the infrastructure, and consequences associated with those failures. However, estimating accidental risks in critical infrastructure involves a substantial effort and costs due to number of variables involved, complexity and lack of information. Therefore, this paper aims to introduce a well trained algorithm for risk assessment using deep learning, which could be capable to deal efficiently with the complexity and uncertainty. The advantage point of the deep learning using near-miss accidents data is that it could be employed in risk assessment as an efficient engineering tool to treat the uncertainty of the risk values in complex environments. The basic idea of using a Near-Miss Deep Learning Approach for Neuro-Fuzzy Risk Assessment in Pipelines is focused in the objective of improve the validity of the risk values learning from near-miss accidents and imitating the human expertise scoring risks and setting tolerance levels. In summary, the method of Deep Learning for Neuro-Fuzzy Risk Assessment involves a regression analysis called group method of data handling (GMDH), which consists in the determination of the optimal configuration of the risk assessment model and its parameters employing polynomial theory.

Keywords: deep learning, risk assessment, neuro fuzzy, pipelines

Procedia PDF Downloads 292

Authors: Heramb Somthankar, Anujoy Chakraborty

Abstract:

This paper presents a classification study on yield prediction in semiconductor production using machine learning approaches. A complicated semiconductor production process is generally monitored continuously by signals acquired from sensors and measurement sites. A monitoring system contains a variety of signals, all of which contain useful information, irrelevant information, and noise. In the case of each signal being considered a feature, "Feature Selection" is used to find the most relevant signals. The open-source UCI SECOM Dataset provides 1567 such samples, out of which 104 fail in quality assurance. Feature extraction and selection are performed on the dataset, and useful signals were considered for further study. Afterward, common machine learning algorithms were employed to predict whether the signal yields pass or fail. The most relevant algorithm is selected for prediction based on the accuracy and loss of the ML model.

Keywords: deep learning, feature extraction, feature selection, machine learning classification algorithms, semiconductor production monitoring, signal processing, time-series analysis

Procedia PDF Downloads 109

Authors: Bikis Muhammed, Sehra Sedigh Sarvestani, Ali R. Hurson, Lasanthi Gamage

Abstract:

Vehicular traffic events have overly complex spatial correlations and temporal interdependencies and are also influenced by environmental events such as weather conditions. To capture these spatial and temporal interdependencies and make more realistic vehicular traffic predictions, graph neural networks (GNN) based traffic prediction models have been extensively utilized due to their capability of capturing non-Euclidean spatial correlation very effectively. However, most of the already existing GNN-based traffic prediction models have some limitations during learning complex and dynamic spatial and temporal patterns due to the following missing factors. First, most GNN-based traffic prediction models have used static distance or sometimes haversine distance mechanisms between spatially separated traffic observations to estimate spatial correlation. Secondly, most GNN-based traffic prediction models have not incorporated environmental events that have a major impact on the normal traffic states. Finally, most of the GNN-based models did not use an attention mechanism to focus on only important traffic observations. The objective of this paper is to study and make real-time vehicular traffic predictions while incorporating the effect of weather conditions. To fill the previously mentioned gaps, our prediction model uses a real-time driving distance between sensors to build a distance matrix or spatial adjacency matrix and capture spatial correlation. In addition, our prediction model considers the effect of six types of weather conditions and has an attention mechanism in both spatial and temporal data aggregation. Our prediction model efficiently captures the spatial and temporal correlation between traffic events, and it relies on the graph attention network (GAT) and Bidirectional bidirectional long short-term memory (Bi-LSTM) plus attention layers and is called GAT-BILSTMA.

Keywords: deep learning, real time prediction, GAT, Bi-LSTM, attention

Procedia PDF Downloads 71

Authors: Jatuphum Ketchatturat

Abstract:

Value-added assessment has been used for developing the model of monitoring and evaluation of student's learning achievement. The steps of model development consist of 1) study and analyisis of the school and the district report system of student achievement and progress, 2) collecting the data of student achievement to develop the value added indicator, 3) developing the system of value-added assessment by participatory action research approach, 4) putting the system of value-added assessment into the educational district of secondary school, 5) determining the quality of the developed system of value-added assessment. The components of the developed model consist of 1) the database of value-added assessment of student's learning achievement, 2) the process of monitoring and evaluation the student's learning achievement, and 3) the reporting system of value-added assessment of student's learning achievement.

Keywords: learning achievement, monitoring and evaluation, value-added assessment

Procedia PDF Downloads 423

Authors: Xiaoyan Dai, Hsieh Yisan

Abstract:

Barcode processing of captured images is a huge challenge, as different shooting conditions can result in different barcode appearances. This paper proposes a deep learning-based barcode detection using synthetic-to-real data augmentation. We first augment barcodes themselves; we then augment images containing the barcodes to generate a large variety of data that is close to the actual shooting environments. Comparisons with previous works and evaluations with our original data show that this approach achieves state-of-the-art performance in various real images. In addition, the system uses hybrid resolution for barcode “scan” and is applicable to real-time applications.

Keywords: barcode detection, data augmentation, deep learning, image-based processing

Procedia PDF Downloads 168

Authors: Jan Stodt, Christoph Reich

Abstract:

The usage of machine learning models for prediction is growing rapidly and proof that the intended requirements are met is essential. Audits are a proven method to determine whether requirements or guidelines are met. However, machine learning models have intrinsic characteristics, such as the quality of training data, that make it difficult to demonstrate the required behavior and make audits more challenging. This paper describes an ML audit framework that evaluates and reviews the risks of machine learning applications, the quality of the training data, and the machine learning model. We evaluate and demonstrate the functionality of the proposed framework by auditing an steel plate fault prediction model.

Keywords: audit, machine learning, assessment, metrics

Procedia PDF Downloads 271

Authors: Aisultan Shoiynbek, Darkhan Kuanyshbay, Paulo Menezes, Akbayan Bekarystankyzy, Assylbek Mukhametzhanov, Temirlan Shoiynbek

Abstract:

Speech emotion recognition (SER) has received increasing research interest in recent years. It is a common practice to utilize emotional speech collected under controlled conditions recorded by actors imitating and artificially producing emotions in front of a microphone. There are four issues related to that approach: emotions are not natural, meaning that machines are learning to recognize fake emotions; emotions are very limited in quantity and poor in variety of speaking; there is some language dependency in SER; consequently, each time researchers want to start work with SER, they need to find a good emotional database in their language. This paper proposes an approach to create an automatic tool for speech emotion extraction based on facial emotion recognition and describes the sequence of actions involved in the proposed approach. One of the first objectives in the sequence of actions is the speech detection issue. The paper provides a detailed description of the speech detection model based on a fully connected deep neural network for Kazakh and Russian. Despite the high results in speech detection for Kazakh and Russian, the described process is suitable for any language. To investigate the working capacity of the developed model, an analysis of speech detection and extraction from real tasks has been performed.

Keywords: deep neural networks, speech detection, speech emotion recognition, Mel-frequency cepstrum coefficients, collecting speech emotion corpus, collecting speech emotion dataset, Kazakh speech dataset

Procedia PDF Downloads 26

Authors: Hyeongbok Kim, Lingling Zhao, Xiaohong Su

Abstract:

Deep reinforcement learning has been applied to address various problems in robotics, such as autonomous driving and unmanned aerial vehicle. However, because of the sparse reward penalty for a collision with obstacles during the navigation mission, the agent fails to learn the optimal policy or requires a long time for convergence. Therefore, using obstacles and enemy agents, in this paper, we present a curriculum-based boost learning method to effectively train compound skills during multi-agent reinforcement learning. First, to enable the agents to solve challenging tasks, we gradually increased learning difficulties by adjusting reward shaping instead of constructing different learning environments. Then, in a benchmark environment with static obstacles and moving enemy agents, the experimental results showed that the proposed curriculum learning strategy enhanced cooperative navigation and compound collision avoidance skills in uncertain environments while improving learning efficiency.

Keywords: curriculum learning, hard exploration, multi-agent reinforcement learning, robotic navigation, sparse reward

Procedia PDF Downloads 92

Authors: Burak Oksuzoglu, Savas Yildirim, Ferhat Kutlu

Abstract:

Opinion mining (OM) is one of the natural language processing (NLP) problems to determine the polarity of opinions, mostly represented on a positive-neutral-negative axis. The data for OM is usually collected from various social media platforms. In an era where social media has considerable control over companies’ futures, it’s worth understanding social media and taking actions accordingly. OM comes to the fore here as the scale of the discussion about companies increases, and it becomes unfeasible to gauge opinion on individual levels. Thus, the companies opt to automize this process by applying machine learning (ML) approaches to their data. For the last two decades, OM or sentiment analysis (SA) has been mainly performed by applying ML classification algorithms such as support vector machines (SVM) and Naïve Bayes to a bag of n-gram representations of textual data. With the advent of deep learning and its apparent success in NLP, traditional methods have become obsolete. Transfer learning paradigm that has been commonly used in computer vision (CV) problems started to shape NLP approaches and language models (LM) lately. This gave a sudden rise to the usage of the pretrained language model (PTM), which contains language representations that are obtained by training it on the large datasets using self-supervised learning objectives. The PTMs are further fine-tuned by a specialized downstream task dataset to produce efficient models for various NLP tasks such as OM, NER (Named-Entity Recognition), Question Answering (QA), and so forth. In this study, the traditional and modern NLP approaches have been evaluated for OM by using a sizable corpus belonging to a large private company containing about 76,000 comments in Turkish: SVM with a bag of n-grams, and two chosen pre-trained models, multilingual universal sentence encoder (MUSE) and bidirectional encoder representations from transformers (BERT). The MUSE model is a multilingual model that supports 16 languages, including Turkish, and it is based on convolutional neural networks. The BERT is a monolingual model in our case and transformers-based neural networks. It uses a masked language model and next sentence prediction tasks that allow the bidirectional training of the transformers. During the training phase of the architecture, pre-processing operations such as morphological parsing, stemming, and spelling correction was not used since the experiments showed that their contribution to the model performance was found insignificant even though Turkish is a highly agglutinative and inflective language. The results show that usage of deep learning methods with pre-trained models and fine-tuning achieve about 11% improvement over SVM for OM. The BERT model achieved around 94% prediction accuracy while the MUSE model achieved around 88% and SVM did around 83%. The MUSE multilingual model shows better results than SVM, but it still performs worse than the monolingual BERT model.

Keywords: BERT, MUSE, opinion mining, pretrained language model, SVM, Turkish

Procedia PDF Downloads 146

Authors: Maroonisha Maharajh

Abstract:

Based on the Black Swan and Disruptive Innovation Theories, this study proposes an educational game based learning model within the context of the traditional classroom learning environment. In the proposed model, the perceived e-learning component is decomposed into accessibility, perceived quality and perceived usability within the traditional rural classroom environment. A sample of 92 respondents took part in this study. The results suggest that users’ continuance intention is determined by both economic and grassroots internet accessibility, which in turn is jointly determined by perceived usefulness, information quality, service quality, system quality, perceived ease of use and cognitive absorption of learning.

Keywords: blended learning, flipped classroom, e-learning, gaming

Procedia PDF Downloads 254

Authors: Xiaoqi Jia, Qing Zhu, Hao Zhang, Su Yang

Abstract:

A gender recognition system is able to tell the gender of the given person through a few of frontal facial images. An effective gender recognition approach enables to improve the performance of many other applications, including security monitoring, human-computer interaction, image or video retrieval and so on. In this paper, we present an effective method for gender classification task in frontal facial images based on deep belief networks (DBNs), which can pre-train model and improve accuracy a little bit. Our experiments have shown that the pre-training method with DBNs for gender classification task is feasible and achieves a little improvement of accuracy on FERET and CAS-PEAL-R1 facial datasets.

Keywords: gender recognition, beep belief net-works, semi-supervised learning, greedy-layer wise RBMs

Procedia PDF Downloads 452

Authors: Yifan Hou, Haibo Liu, Le Jiang, Wandong Su, Binqing Wang

Abstract:

Road defect detection is crucial for modern urban management and infrastructure maintenance. Traditional road defect detection methods mostly rely on manual labor, which is not only inefficient but also difficult to ensure their reliability. However, existing deep learning-based road defect detection models have poor detection performance in complex environments and lack robustness to multi-scale targets. To address this challenge, this paper proposes a distinct detection framework based on the one stage algorithm network structure. This article designs a deep feature extraction network based on RCSDarknet, which applies channel shuffling to enhance information fusion between tensors. Through repeated stacking of RCS modules, the information flow between different channels of adjacent layer features is enhanced to improve the model's ability to capture target spatial features. In addition, a multi-scale feature fusion mechanism with weighted dual flow paths was adopted to fuse spatial features of different scales, thereby further improving the detection performance of the model at different scales. To validate the performance of the proposed algorithm, we tested it using the RDD2022 dataset. The experimental results show that the enhancement algorithm achieved 84.14% mAP, which is 1.06% higher than the currently advanced YOLOv8 algorithm. Through visualization analysis of the results, it can also be seen that our proposed algorithm has good performance in detecting targets of different scales in complex scenes. The above experimental results demonstrate the effectiveness and superiority of the proposed algorithm, providing valuable insights for advancing real-time road defect detection methods.

Keywords: roads, defect detection, visualization, deep learning

Procedia PDF Downloads 6