Commenced in January 2007
Frequency: Monthly
Edition: International
Paper Count: 85

Deep learning Related Abstracts

85 Forward Conditional Restricted Boltzmann Machines for the Generation of Music

Authors: Johan Loeckx, Joeri Bultheel


Recently, the application of deep learning to music has gained popularity. Its true potential, however, has been largely unexplored. In this paper, a new idea for representing the dynamic behavior of music is proposed. A ”forward” conditional RBM takes into account not only preceding but also future samples during training. Though this may sound controversial at first sight, it will be shown that it makes sense from a musical and neuro-cognitive perspective. The model is applied to reconstruct music based upon the first notes and to improvise in the musical style of a composer. Different to expectations, reconstruction accuracy with respect to a regular CRBM with the same order, was not significantly improved. More research is needed to test the performance on unseen data.

Keywords: Deep learning, restricted boltzmann machine, music generation, conditional restricted boltzmann machine (CRBM)

Procedia PDF Downloads 375
84 Health Trajectory Clustering Using Deep Belief Networks

Authors: Farshid Hajati, Federico Girosi, Shima Ghassempour


We present a Deep Belief Network (DBN) method for clustering health trajectories. Deep Belief Network (DBN) is a deep architecture that consists of a stack of Restricted Boltzmann Machines (RBM). In a deep architecture, each layer learns more complex features than the past layers. The proposed method depends on DBN in clustering without using back propagation learning algorithm. The proposed DBN has a better a performance compared to the deep neural network due the initialization of the connecting weights. We use Contrastive Divergence (CD) method for training the RBMs which increases the performance of the network. The performance of the proposed method is evaluated extensively on the Health and Retirement Study (HRS) database. The University of Michigan Health and Retirement Study (HRS) is a nationally representative longitudinal study that has surveyed more than 27,000 elderly and near-elderly Americans since its inception in 1992. Participants are interviewed every two years and they collect data on physical and mental health, insurance coverage, financial status, family support systems, labor market status, and retirement planning. The dataset is publicly available and we use the RAND HRS version L, which is easy to use and cleaned up version of the data. The size of sample data set is 268 and the length of the trajectories is equal to 10. The trajectories do not stop when the patient dies and represent 10 different interviews of live patients. Compared to the state-of-the-art benchmarks, the experimental results show the effectiveness and superiority of the proposed method in clustering health trajectories.

Keywords: Clustering, Deep learning, DBN, health trajectory

Procedia PDF Downloads 234
83 DNpro: A Deep Learning Network Approach to Predicting Protein Stability Changes Induced by Single-Site Mutations

Authors: Xiao Zhou, Jianlin Cheng


A single amino acid mutation can have a significant impact on the stability of protein structure. Thus, the prediction of protein stability change induced by single site mutations is critical and useful for studying protein function and structure. Here, we presented a deep learning network with the dropout technique for predicting protein stability changes upon single amino acid substitution. While using only protein sequence as input, the overall prediction accuracy of the method on a standard benchmark is >85%, which is higher than existing sequence-based methods and is comparable to the methods that use not only protein sequence but also tertiary structure, pH value and temperature. The results demonstrate that deep learning is a promising technique for protein stability prediction. The good performance of this sequence-based method makes it a valuable tool for predicting the impact of mutations on most proteins whose experimental structures are not available. Both the downloadable software package and the user-friendly web server (DNpro) that implement the method for predicting protein stability changes induced by amino acid mutations are freely available for the community to use.

Keywords: Bioinformatics, Biological Data Mining, Deep learning, protein stability prediction

Procedia PDF Downloads 261
82 3D Plant Growth Measurement System Using Deep Learning Technology

Authors: Kazuaki Shiraishi, Narumitsu Asai, Tsukasa Kitahara, Sosuke Mieno, Takaharu Kameoka


The purpose of this research is to facilitate productivity advances in agriculture. To accomplish this, we developed an automatic three-dimensional (3D) recording system for growth of field crops that consists of a number of inexpensive modules: a very low-cost stereo camera, a couple of ZigBee wireless modules, a Raspberry Pi single-board computer, and a third generation (3G) wireless communication module. Our system uses an inexpensive Web stereo camera in order to keep total costs low. However, inexpensive video cameras record low-resolution images that are very noisy. Accordingly, in order to resolve these problems, we adopted a deep learning method. Based on the results of extended period of time operation test conducted without the use of an external power supply, we found that by using Super-Resolution Convolutional Neural Network method, our system could achieve a balance between the competing goals of low-cost and superior performance. Our experimental results showed the effectiveness of our system.

Keywords: Image Processing, Deep learning, automatic recording, stereo camera

Procedia PDF Downloads 135
81 Embedded Visual Perception for Autonomous Agricultural Machines Using Lightweight Convolutional Neural Networks

Authors: Henrik Karstoft, René A. Sørensen, Søren Skovsen, Peter Christiansen


Autonomous agricultural machines act in stochastic surroundings and therefore, must be able to perceive the surroundings in real time. This perception can be achieved using image sensors combined with advanced machine learning, in particular Deep Learning. Deep convolutional neural networks excel in labeling and perceiving color images and since the cost of high-quality RGB-cameras is low, the hardware cost of good perception depends heavily on memory and computation power. This paper investigates the possibility of designing lightweight convolutional neural networks for semantic segmentation (pixel wise classification) with reduced hardware requirements, to allow for embedded usage in autonomous agricultural machines. Using compression techniques, a lightweight convolutional neural network is designed to perform real-time semantic segmentation on an embedded platform. The network is trained on two large datasets, ImageNet and Pascal Context, to recognize up to 400 individual classes. The 400 classes are remapped into agricultural superclasses (e.g. human, animal, sky, road, field, shelterbelt and obstacle) and the ability to provide accurate real-time perception of agricultural surroundings is studied. The network is applied to the case of autonomous grass mowing using the NVIDIA Tegra X1 embedded platform. Feeding case-specific images to the network results in a fully segmented map of the superclasses in the image. As the network is still being designed and optimized, only a qualitative analysis of the method is complete at the abstract submission deadline. Proceeding this deadline, the finalized design is quantitatively evaluated on 20 annotated grass mowing images. Lightweight convolutional neural networks for semantic segmentation can be implemented on an embedded platform and show competitive performance with regards to accuracy and speed. It is feasible to provide cost-efficient perceptive capabilities related to semantic segmentation for autonomous agricultural machines.

Keywords: Safety, Deep learning, Visual Perception, autonomous agricultural machines

Procedia PDF Downloads 273
80 A Study of Various Ontology Learning Systems from Text and a Look into Future

Authors: Fatima Al-Aswadi, Chan Yong


With the large volume of unstructured data that increases day by day on the web, the motivation of representing the knowledge in this data in the machine processable form is increased. Ontology is one of the major cornerstones of representing the information in a more meaningful way on the semantic Web. The goal of Ontology learning from text is to elicit and represent domain knowledge in the machine readable form. This paper aims to give a follow-up review on the ontology learning systems from text and some of their defects. Furthermore, it discusses how far the ontology learning process will enhance in the future.

Keywords: Semantic Web, Deep learning, concept discovery, ontology learning, semantic relation

Procedia PDF Downloads 276
79 Artificial Intelligence in Bioscience: The Next Frontier

Authors: Parthiban Srinivasan


With recent advances in computational power and access to enough data in biosciences, artificial intelligence methods are increasingly being used in drug discovery research. These methods are essentially a series of advanced statistics based exercises that review the past to indicate the likely future. Our goal is to develop a model that accurately predicts biological activity and toxicity parameters for novel compounds. We have compiled a robust library of over 150,000 chemical compounds with different pharmacological properties from literature and public domain databases. The compounds are stored in simplified molecular-input line-entry system (SMILES), a commonly used text encoding for organic molecules. We utilize an automated process to generate an array of numerical descriptors (features) for each molecule. Redundant and irrelevant descriptors are eliminated iteratively. Our prediction engine is based on a portfolio of machine learning algorithms. We found Random Forest algorithm to be a better choice for this analysis. We captured non-linear relationship in the data and formed a prediction model with reasonable accuracy by averaging across a large number of randomized decision trees. Our next step is to apply deep neural network (DNN) algorithm to predict the biological activity and toxicity properties. We expect the DNN algorithm to give better results and improve the accuracy of the prediction. This presentation will review all these prominent machine learning and deep learning methods, our implementation protocols and discuss these techniques for their usefulness in biomedical and health informatics.

Keywords: Machine Learning, Health Informatics, Drug discovery, Deep learning, toxicity prediction

Procedia PDF Downloads 256
78 Near-Miss Deep Learning Approach for Neuro-Fuzzy Risk Assessment in Pipelines

Authors: Alexander Guzman Urbina, Atsushi Aoyama


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: Risk Assessment, Deep learning, Pipelines, neuro fuzzy

Procedia PDF Downloads 172
77 Genetic Algorithm Based Deep Learning Parameters Tuning for Robot Object Recognition and Grasping

Authors: Delowar Hossain, Genci Capi


This paper concerns with the problem of deep learning parameters tuning using a genetic algorithm (GA) in order to improve the performance of deep learning (DL) method. We present a GA based DL method for robot object recognition and grasping. GA is used to optimize the DL parameters in learning procedure in term of the fitness function that is good enough. After finishing the evolution process, we receive the optimal number of DL parameters. To evaluate the performance of our method, we consider the object recognition and robot grasping tasks. Experimental results show that our method is efficient for robot object recognition and grasping.

Keywords: Object recognition, Genetic Algorithm, Deep learning, robot grasping

Procedia PDF Downloads 204
76 Augmented Reality Sandbox and Constructivist Approach for Geoscience Teaching and Learning

Authors: Muhammad Nawaz, Farha Sattar, Sandeep N. Kundu


Augmented reality sandbox adds new dimensions to education and learning process. It can be a core component of geoscience teaching and learning to understand the geographic contexts and landform processes. Augmented reality sandbox is a useful tool not only to create an interactive learning environment through spatial visualization but also it can provide an active learning experience to students and enhances the cognition process of learning. Augmented reality sandbox can be used as an interactive learning tool to teach geomorphic and landform processes. This article explains the augmented reality sandbox and the constructivism approach for geoscience teaching and learning, and endeavours to explore the ways to teach the geographic processes using the three-dimensional digital environment for the deep learning of the geoscience concepts interactively.

Keywords: Geoscience, Deep learning, constructivism, augmented reality sandbox

Procedia PDF Downloads 241
75 A Review of Machine Learning for Big Data

Authors: Devatha Kalyan Kumar, Aravindraj D., Sadathulla A.


Big data are now rapidly expanding in all engineering and science and many other domains. The potential of large or massive data is undoubtedly significant, make sense to require new ways of thinking and learning techniques to address the various big data challenges. Machine learning is continuously unleashing its power in a wide range of applications. In this paper, the latest advances and advancements in the researches on machine learning for big data processing. First, the machine learning techniques methods in recent studies, such as deep learning, representation learning, transfer learning, active learning and distributed and parallel learning. Then focus on the challenges and possible solutions of machine learning for big data.

Keywords: Machine Learning, Big Data, Active Learning, Deep learning

Procedia PDF Downloads 225
74 Deep Learning for Recommender System: Principles, Methods and Evaluation

Authors: Basiliyos Tilahun Betru, Charles Awono Onana, Bernabe Batchakui


Recommender systems have become increasingly popular in recent years, and are utilized in numerous areas. Nowadays many web services provide several information for users and recommender systems have been developed as critical element of these web applications to predict choice of preference and provide significant recommendations. With the help of the advantage of deep learning in modeling different types of data and due to the dynamic change of user preference, building a deep model can better understand users demand and further improve quality of recommendation. In this paper, deep neural network models for recommender system are evaluated. Most of deep neural network models in recommender system focus on the classical collaborative filtering user-item setting. Deep learning models demonstrated high level features of complex data can be learned instead of using metadata which can significantly improve accuracy of recommendation. Even though deep learning poses a great impact in various areas, applying the model to a recommender system have not been fully exploited and still a lot of improvements can be done both in collaborative and content-based approach while considering different contextual factors.

Keywords: Decision Making, Big Data, Recommender System, Deep learning

Procedia PDF Downloads 183
73 Deep Learning Based on Image Decomposition for Restoration of Intrinsic Representation

Authors: Hyohun Kim, Dongwha Shin, Yeonseok Kim, Ji-Su Ahn, Dongeun Choi, Byung-Woo Hong, Kensuke Nakamura


Artefacts are commonly encountered in the imaging process of clinical computed tomography (CT) where the artefact refers to any systematic discrepancy between the reconstructed observation and the true attenuation coefficient of the object. It is known that CT images are inherently more prone to artefacts due to its image formation process where a large number of independent detectors are involved, and they are assumed to yield consistent measurements. There are a number of different artefact types including noise, beam hardening, scatter, pseudo-enhancement, motion, helical, ring, and metal artefacts, which cause serious difficulties in reading images. Thus, it is desired to remove nuisance factors from the degraded image leaving the fundamental intrinsic information that can provide better interpretation of the anatomical and pathological characteristics. However, it is considered as a difficult task due to the high dimensionality and variability of data to be recovered, which naturally motivates the use of machine learning techniques. We propose an image restoration algorithm based on the deep neural network framework where the denoising auto-encoders are stacked building multiple layers. The denoising auto-encoder is a variant of a classical auto-encoder that takes an input data and maps it to a hidden representation through a deterministic mapping using a non-linear activation function. The latent representation is then mapped back into a reconstruction the size of which is the same as the size of the input data. The reconstruction error can be measured by the traditional squared error assuming the residual follows a normal distribution. In addition to the designed loss function, an effective regularization scheme using residual-driven dropout determined based on the gradient at each layer. The optimal weights are computed by the classical stochastic gradient descent algorithm combined with the back-propagation algorithm. In our algorithm, we initially decompose an input image into its intrinsic representation and the nuisance factors including artefacts based on the classical Total Variation problem that can be efficiently optimized by the convex optimization algorithm such as primal-dual method. The intrinsic forms of the input images are provided to the deep denosing auto-encoders with their original forms in the training phase. In the testing phase, a given image is first decomposed into the intrinsic form and then provided to the trained network to obtain its reconstruction. We apply our algorithm to the restoration of the corrupted CT images by the artefacts. It is shown that our algorithm improves the readability and enhances the anatomical and pathological properties of the object. The quantitative evaluation is performed in terms of the PSNR, and the qualitative evaluation provides significant improvement in reading images despite degrading artefacts. The experimental results indicate the potential of our algorithm as a prior solution to the image interpretation tasks in a variety of medical imaging applications. This work was supported by the MISP(Ministry of Science and ICT), Korea, under the National Program for Excellence in SW (20170001000011001) supervised by the IITP(Institute for Information and Communications Technology Promotion).

Keywords: Deep learning, Noise Reduction, auto-encoder neural network, CT image artefact, intrinsic image representation, total variation

Procedia PDF Downloads 60
72 An Intelligent Baby Care System Based on IoT and Deep Learning Techniques

Authors: Chinlun Lai, Lunjyh Jiang


Due to the heavy burden and pressure of caring for infants, an integrated automatic baby watching system based on IoT smart sensing and deep learning machine vision techniques is proposed in this paper. By monitoring infant body conditions such as heartbeat, breathing, body temperature, sleeping posture, as well as the surrounding conditions such as dangerous/sharp objects, light, noise, humidity and temperature, the proposed system can analyze and predict the obvious/potential dangerous conditions according to observed data and then adopt suitable actions in real time to protect the infant from harm. Thus, reducing the burden of the caregiver and improving safety efficiency of the caring work. The experimental results show that the proposed system works successfully for the infant care work and thus can be implemented in various life fields practically.

Keywords: Internet of Things, Machine Vision, Deep learning, baby care system

Procedia PDF Downloads 102
71 Convolutional Neural Network Based on Random Kernels for Analyzing Visual Imagery

Authors: Hyohun Kim, Dongwha Shin, Yeonseok Kim, Ji-Su Ahn, Byung-Woo Hong, Kensuke Nakamura, Ja-Keoung Koo


The machine learning techniques based on a convolutional neural network (CNN) have been actively developed and successfully applied to a variety of image analysis tasks including reconstruction, noise reduction, resolution enhancement, segmentation, motion estimation, object recognition. The classical visual information processing that ranges from low level tasks to high level ones has been widely developed in the deep learning framework. It is generally considered as a challenging problem to derive visual interpretation from high dimensional imagery data. A CNN is a class of feed-forward artificial neural network that usually consists of deep layers the connections of which are established by a series of non-linear operations. The CNN architecture is known to be shift invariant due to its shared weights and translation invariance characteristics. However, it is often computationally intractable to optimize the network in particular with a large number of convolution layers due to a large number of unknowns to be optimized with respect to the training set that is generally required to be large enough to effectively generalize the model under consideration. It is also necessary to limit the size of convolution kernels due to the computational expense despite of the recent development of effective parallel processing machinery, which leads to the use of the constantly small size of the convolution kernels throughout the deep CNN architecture. However, it is often desired to consider different scales in the analysis of visual features at different layers in the network. Thus, we propose a CNN model where different sizes of the convolution kernels are applied at each layer based on the random projection. We apply random filters with varying sizes and associate the filter responses with scalar weights that correspond to the standard deviation of the random filters. We are allowed to use large number of random filters with the cost of one scalar unknown for each filter. The computational cost in the back-propagation procedure does not increase with the larger size of the filters even though the additional computational cost is required in the computation of convolution in the feed-forward procedure. The use of random kernels with varying sizes allows to effectively analyze image features at multiple scales leading to a better generalization. The robustness and effectiveness of the proposed CNN based on random kernels are demonstrated by numerical experiments where the quantitative comparison of the well-known CNN architectures and our models that simply replace the convolution kernels with the random filters is performed. The experimental results indicate that our model achieves better performance with less number of unknown weights. The proposed algorithm has a high potential in the application of a variety of visual tasks based on the CNN framework. Acknowledgement—This work was supported by the MISP (Ministry of Science and ICT), Korea, under the National Program for Excellence in SW (20170001000011001) supervised by IITP, and NRF-2014R1A2A1A11051941, NRF2017R1A2B4006023.

Keywords: Object recognition, Deep learning, dimensionality reduction, random projection, convolutional neural network, random kernel

Procedia PDF Downloads 87
70 Using Deep Learning Real-Time Object Detection Convolution Neural Networks for Fast Fruit Recognition in the Tree

Authors: K. Bresilla, L. Manfrini, B. Morandi, A. Boini, G. Perulli, L. C. Grappadelli


Image/video processing for fruit in the tree using hard-coded feature extraction algorithms have shown high accuracy during recent years. While accurate, these approaches even with high-end hardware are computationally intensive and too slow for real-time systems. This paper details the use of deep convolution neural networks (CNNs), specifically an algorithm (YOLO - You Only Look Once) with 24+2 convolution layers. Using deep-learning techniques eliminated the need for hard-code specific features for specific fruit shapes, color and/or other attributes. This CNN is trained on more than 5000 images of apple and pear fruits on 960 cores GPU (Graphical Processing Unit). Testing set showed an accuracy of 90%. After this, trained data were transferred to an embedded device (Raspberry Pi gen.3) with camera for more portability. Based on correlation between number of visible fruits or detected fruits on one frame and the real number of fruits on one tree, a model was created to accommodate this error rate. Speed of processing and detection of the whole platform was higher than 40 frames per second. This speed is fast enough for any grasping/harvesting robotic arm or other real-time applications.

Keywords: Artificial Intelligence, Computer Vision, Precision Agriculture, Deep learning, fruit recognition, harvesting robot

Procedia PDF Downloads 259
69 Stock Market Prediction Using Convolutional Neural Network That Learns from a Graph

Authors: Hyunchul Ahn, Kee-Young Kwahk, Mo-Se Lee, Cheol-Hwi Ahn


Over the past decade, deep learning has been in spotlight among various machine learning algorithms. In particular, CNN (Convolutional Neural Network), which is known as effective solution for recognizing and classifying images, has been popularly applied to classification and prediction problems in various fields. In this study, we try to apply CNN to stock market prediction, one of the most challenging tasks in the machine learning research. In specific, we propose to apply CNN as the binary classifier that predicts stock market direction (up or down) by using a graph as its input. That is, our proposal is to build a machine learning algorithm that mimics a person who looks at the graph and predicts whether the trend will go up or down. Our proposed model consists of four steps. In the first step, it divides the dataset into 5 days, 10 days, 15 days, and 20 days. And then, it creates graphs for each interval in step 2. In the next step, CNN classifiers are trained using the graphs generated in the previous step. In step 4, it optimizes the hyper parameters of the trained model by using the validation dataset. To validate our model, we will apply it to the prediction of KOSPI200 for 1,986 days in eight years (from 2009 to 2016). The experimental dataset will include 14 technical indicators such as CCI, Momentum, ROC and daily closing price of KOSPI200 of Korean stock market.

Keywords: Deep learning, stock market prediction, convolutional neural network, Korean stock market

Procedia PDF Downloads 305
68 R-Killer: An Email-Based Ransomware Protection Tool

Authors: B. Lokuketagoda, M. Weerakoon, U. Madushan, A. N. Senaratne, K. Y. Abeywardena


Ransomware has become a common threat in past few years and the recent threat reports show an increase of growth in Ransomware infections. Researchers have identified different variants of Ransomware families since 2015. Lack of knowledge of the user about the threat is a major concern. Ransomware detection methodologies are still growing through the industry. Email is the easiest method to send Ransomware to its victims. Uninformed users tend to click on links and attachments without much consideration assuming the emails are genuine. As a solution to this in this paper R-Killer Ransomware detection tool is introduced. Tool can be integrated with existing email services. The core detection Engine (CDE) discussed in the paper focuses on separating suspicious samples from emails and handling them until a decision is made regarding the suspicious mail. It has the capability of preventing execution of identified ransomware processes. On the other hand, Sandboxing and URL analyzing system has the capability of communication with public threat intelligence services to gather known threat intelligence. The R-Killer has its own mechanism developed in its Proactive Monitoring System (PMS) which can monitor the processes created by downloaded email attachments and identify potential Ransomware activities. R-killer is capable of gathering threat intelligence without exposing the user’s data to public threat intelligence services, hence protecting the confidentiality of user data.

Keywords: Deep learning, Ransomware, Recurrent Neural Networks, email, core detection engine

Procedia PDF Downloads 51
67 Deep Learning-Based Object Detection on Low Quality Images: A Case Study of Real-Time Traffic Monitoring

Authors: Jean-Francois Rajotte, Martin Sotir, Frank Gouineau


The installation and management of traffic monitoring devices can be costly from both a financial and resource point of view. It is therefore important to take advantage of in-place infrastructures to extract the most information. Here we show how low-quality urban road traffic images from cameras already available in many cities (such as Montreal, Vancouver, and Toronto) can be used to estimate traffic flow. To this end, we use a pre-trained neural network, developed for object detection, to count vehicles within images. We then compare the results with human annotations gathered through crowdsourcing campaigns. We use this comparison to assess performance and calibrate the neural network annotations. As a use case, we consider six months of continuous monitoring over hundreds of cameras installed in the city of Montreal. We compare the results with city-provided manual traffic counting performed in similar conditions at the same location. The good performance of our system allows us to consider applications which can monitor the traffic conditions in near real-time, making the counting usable for traffic-related services. Furthermore, the resulting annotations pave the way for building a historical vehicle counting dataset to be used for analysing the impact of road traffic on many city-related issues, such as urban planning, security, and pollution.

Keywords: Artificial Intelligence, Vehicles, Real-Time Systems, Deep learning, Roads, Traffic Monitoring, image annotation

Procedia PDF Downloads 79
66 Hierarchical Tree Long Short-Term Memory for Sentence Representations

Authors: Bo Xu, Xiuying Wang, Changliang Li


A fixed-length feature vector is required for many machine learning algorithms in NLP field. Word embeddings have been very successful at learning lexical information. However, they cannot capture the compositional meaning of sentences, which prevents them from a deeper understanding of language. In this paper, we introduce a novel hierarchical tree long short-term memory (HTLSTM) model that learns vector representations for sentences of arbitrary syntactic type and length. We propose to split one sentence into three hierarchies: short phrase, long phrase and full sentence level. The HTLSTM model gives our algorithm the potential to fully consider the hierarchical information and long-term dependencies of language. We design the experiments on both English and Chinese corpus to evaluate our model on sentiment analysis task. And the results show that our model outperforms several existing state of the art approaches significantly.

Keywords: sentiment analysis, Deep learning, hierarchical tree long short-term memory, sentence representation

Procedia PDF Downloads 210
65 Prediction on Housing Price Based on Deep Learning

Authors: Yan Wang, Li Yu, Chenlu Jiao, Hongrun Xin, Kaiyang Wang


In order to study the impact of various factors on the housing price, we propose to build different prediction models based on deep learning to determine the existing data of the real estate in order to more accurately predict the housing price or its changing trend in the future. Considering that the factors which affect the housing price vary widely, the proposed prediction models include two categories. The first one is based on multiple characteristic factors of the real estate. We built Convolution Neural Network (CNN) prediction model and Long Short-Term Memory (LSTM) neural network prediction model based on deep learning, and logical regression model was implemented to make a comparison between these three models. Another prediction model is time series model. Based on deep learning, we proposed an LSTM-1 model purely regard to time series, then implementing and comparing the LSTM model and the Auto-Regressive and Moving Average (ARMA) model. In this paper, comprehensive study of the second-hand housing price in Beijing has been conducted from three aspects: crawling and analyzing, housing price predicting, and the result comparing. Ultimately the best model program was produced, which is of great significance to evaluation and prediction of the housing price in the real estate industry.

Keywords: Deep learning, convolutional neural network, LSTM, housing prediction

Procedia PDF Downloads 147
64 Use Cloud-Based Watson Deep Learning Platform to Train Models Faster and More Accurate

Authors: Susan Diamond


Machine Learning workloads have traditionally been run in high-performance computing (HPC) environments, where users log in to dedicated machines and utilize the attached GPUs to run training jobs on huge datasets. Training of large neural network models is very resource intensive, and even after exploiting parallelism and accelerators such as GPUs, a single training job can still take days. Consequently, the cost of hardware is a barrier to entry. Even when upfront cost is not a concern, the lead time to set up such an HPC environment takes months from acquiring hardware to set up the hardware with the right set of firmware, software installed and configured. Furthermore, scalability is hard to achieve in a rigid traditional lab environment. Therefore, it is slow to react to the dynamic change in the artificial intelligent industry. Watson Deep Learning as a service, a cloud-based deep learning platform that mitigates the long lead time and high upfront investment in hardware. It enables robust and scalable sharing of resources among the teams in an organization. It is designed for on-demand cloud environments. Providing a similar user experience in a multi-tenant cloud environment comes with its own unique challenges regarding fault tolerance, performance, and security. Watson Deep Learning as a service tackles these challenges and present a deep learning stack for the cloud environments in a secure, scalable and fault-tolerant manner. It supports a wide range of deep-learning frameworks such as Tensorflow, PyTorch, Caffe, Torch, Theano, and MXNet etc. These frameworks reduce the effort and skillset required to design, train, and use deep learning models. Deep Learning as a service is used at IBM by AI researchers in areas including machine translation, computer vision, and healthcare. 

Keywords: Machine Learning, Deep learning, Cognitive Computing, model training

Procedia PDF Downloads 81
63 Application of Deep Learning in Colorization of LiDAR-Derived Intensity Images

Authors: Edgardo V. Gubatanga Jr., Mark Joshua Salvacion


Most aerial LiDAR systems have accompanying aerial cameras in order to capture not only the terrain of the surveyed area but also its true-color appearance. However, the presence of atmospheric clouds, poor lighting conditions, and aerial camera problems during an aerial survey may cause absence of aerial photographs. These leave areas having terrain information but lacking aerial photographs. Intensity images can be derived from LiDAR data but they are only grayscale images. A deep learning model is developed to create a complex function in a form of a deep neural network relating the pixel values of LiDAR-derived intensity images and true-color images. This complex function can then be used to predict the true-color images of a certain area using intensity images from LiDAR data. The predicted true-color images do not necessarily need to be accurate compared to the real world. They are only intended to look realistic so that they can be used as base maps.

Keywords: Deep learning, aerial LiDAR, colorization, intensity images

Procedia PDF Downloads 19
62 Automatic Classification of Periodic Heart Sounds Using Convolutional Neural Network

Authors: Jia Xin Low, Keng Wah Choo


This paper presents an automatic normal and abnormal heart sound classification model developed based on deep learning algorithm. MITHSDB heart sounds datasets obtained from the 2016 PhysioNet/Computing in Cardiology Challenge database were used in this research with the assumption that the electrocardiograms (ECG) were recorded simultaneously with the heart sounds (phonocardiogram, PCG). The PCG time series are segmented per heart beat, and each sub-segment is converted to form a square intensity matrix, and classified using convolutional neural network (CNN) models. This approach removes the need to provide classification features for the supervised machine learning algorithm. Instead, the features are determined automatically through training, from the time series provided. The result proves that the prediction model is able to provide reasonable and comparable classification accuracy despite simple implementation. This approach can be used for real-time classification of heart sounds in Internet of Medical Things (IoMT), e.g. remote monitoring applications of PCG signal.

Keywords: Deep learning, discrete wavelet transform, convolutional neural network, heart sound classification

Procedia PDF Downloads 198
61 Estimating Algae Concentration Based on Deep Learning from Satellite Observation in Korea

Authors: Joon Ha Kim, Heewon Jeong, Seongpyo Kim


Over the last few tens of years, the coastal regions of Korea have experienced red tide algal blooms, which are harmful and toxic to both humans and marine organisms due to their potential threat. It was accelerated owing to eutrophication by human activities, certain oceanic processes, and climate change. Previous studies have tried to monitoring and predicting the algae concentration of the ocean with the bio-optical algorithms applied to color images of the satellite. However, the accurate estimation of algal blooms remains problems to challenges because of the complexity of coastal waters. Therefore, this study suggests a new method to identify the concentration of red tide algal bloom from images of geostationary ocean color imager (GOCI) which are representing the water environment of the sea in Korea. The method employed GOCI images, which took the water leaving radiances centered at 443nm, 490nm and 660nm respectively, as well as observed weather data (i.e., humidity, temperature and atmospheric pressure) for the database to apply optical characteristics of algae and train deep learning algorithm. Convolution neural network (CNN) was used to extract the significant features from the images. And then artificial neural network (ANN) was used to estimate the concentration of algae from the extracted features. For training of the deep learning model, backpropagation learning strategy is developed. The established methods were tested and compared with the performances of GOCI data processing system (GDPS), which is based on standard image processing algorithms and optical algorithms. The model had better performance to estimate algae concentration than the GDPS which is impossible to estimate greater than 5mg/m³. Thus, deep learning model trained successfully to assess algae concentration in spite of the complexity of water environment. Furthermore, the results of this system and methodology can be used to improve the performances of remote sensing. Acknowledgement: This work was supported by the 'Climate Technology Development and Application' research project (#K07731) through a grant provided by GIST in 2017.

Keywords: Satellite, Remote Sensing, Deep learning, algae concentration

Procedia PDF Downloads 55
60 Deep Learning Based Road Crack Detection on an Embedded Platform

Authors: Ayhan Küçükmanisa, Oğuzhan Urhan, Nurhak Altın


It is important that highways are in good condition for traffic safety. Road crashes (road cracks, erosion of lane markings, etc.) can cause accidents by affecting driving. Image processing based methods for detecting road cracks are available in the literature. In this paper, a deep learning based road crack detection approach is proposed. YOLO (You Look Only Once) is adopted as core component of the road crack detection approach presented. The YOLO network structure, which is developed for object detection, is trained with road crack images as a new class that is not previously used in YOLO. The performance of the proposed method is compared using different training methods: using randomly generated weights and training their own pre-trained weights (transfer learning). A similar training approach is applied to the simplified version of the YOLO network model (tiny yolo) and the results of the performance are examined. The developed system is able to process 8 fps on NVIDIA Jetson TX1 development kit.

Keywords: Deep learning, embedded platform, real-time processing, road crack detection

Procedia PDF Downloads 182
59 Decoding Kinematic Characteristics of Finger Movement from Electrocorticography Using Classical Methods and Deep Convolutional Neural Networks

Authors: Ksenia Volkova, Artur Petrosyan, Ignatii Dubyshkin, Alexei Ossadtchi


Brain-computer interfaces are a growing research field producing many implementations that find use in different fields and are used for research and practical purposes. Despite the popularity of the implementations using non-invasive neuroimaging methods, radical improvement of the state channel bandwidth and, thus, decoding accuracy is only possible by using invasive techniques. Electrocorticography (ECoG) is a minimally invasive neuroimaging method that provides highly informative brain activity signals, effective analysis of which requires the use of machine learning methods that are able to learn representations of complex patterns. Deep learning is a family of machine learning algorithms that allow learning representations of data with multiple levels of abstraction. This study explores the potential of deep learning approaches for ECoG processing, decoding movement intentions and the perception of proprioceptive information. To obtain synchronous recording of kinematic movement characteristics and corresponding electrical brain activity, a series of experiments were carried out, during which subjects performed finger movements at their own pace. Finger movements were recorded with a three-axis accelerometer, while ECoG was synchronously registered from the electrode strips that were implanted over the contralateral sensorimotor cortex. Then, multichannel ECoG signals were used to track finger movement trajectory characterized by accelerometer signal. This process was carried out both causally and non-causally, using different position of the ECoG data segment with respect to the accelerometer data stream. The recorded data was split into training and testing sets, containing continuous non-overlapping fragments of the multichannel ECoG. A deep convolutional neural network was implemented and trained, using 1-second segments of ECoG data from the training dataset as input. To assess the decoding accuracy, correlation coefficient r between the output of the model and the accelerometer readings was computed. After optimization of hyperparameters and training, the deep learning model allowed reasonably accurate causal decoding of finger movement with correlation coefficient r = 0.8. In contrast, the classical Wiener-filter like approach was able to achieve only 0.56 in the causal decoding mode. In the noncausal case, the traditional approach reached the accuracy of r = 0.69, which may be due to the presence of additional proprioceptive information. This result demonstrates that the deep neural network was able to effectively find a representation of the complex top-down information related to the actual movement rather than proprioception. The sensitivity analysis shows physiologically plausible pictures of the extent to which individual features (channel, wavelet subband) are utilized during the decoding procedure. In conclusion, the results of this study have demonstrated that a combination of a minimally invasive neuroimaging technique such as ECoG and advanced machine learning approaches allows decoding motion with high accuracy. Such setup provides means for control of devices with a large number of degrees of freedom as well as exploratory studies of the complex neural processes underlying movement execution.

Keywords: Deep learning, Brain-Computer Interface, ECoG, movement decoding, sensorimotor cortex

Procedia PDF Downloads 26
58 An Accurate Brain Tumor Segmentation for High Graded Glioma Using Deep Learning

Authors: Asad Ali Safi, Sajeeha Ansar, Sheikh Ziauddin, Ahmad R. Shahid, Faraz Ahsan


Gliomas are most challenging and aggressive type of tumors which appear in different sizes, locations, and scattered boundaries. CNN is most efficient deep learning approach with outstanding capability of solving image analysis problems. A fully automatic deep learning based 2D-CNN model for brain tumor segmentation is presented in this paper. We used small convolution filters (3 x 3) to make architecture deeper. We increased convolutional layers for efficient learning of complex features from large dataset. We achieved better results by pushing convolutional layers up to 16 layers for HGG model. We achieved reliable and accurate results through fine-tuning among dataset and hyper-parameters. Pre-processing of this model includes generation of brain pipeline, intensity normalization, bias correction and data augmentation. We used the BRATS-2015, and Dice Similarity Coefficient (DSC) is used as performance measure for the evaluation of the proposed method. Our method achieved DSC score of 0.81 for complete, 0.79 for core, 0.80 for enhanced tumor regions. However, these results are comparable with methods already implemented 2D CNN architecture.

Keywords: Deep learning, Convolutional Neural Networks, brain tumor segmentation, HGG

Procedia PDF Downloads 55
57 LGG Architecture for Brain Tumor Segmentation Using Convolutional Neural Network

Authors: Asad Ali Safi, Sajeeha Ansar, Sheikh Ziauddin, Ahmad R. Shahid, Faraz Ahsan


The most aggressive form of brain tumor is called glioma. Glioma is kind of tumor that arises from glial tissue of the brain and occurs quite often. A fully automatic 2D-CNN model for brain tumor segmentation is presented in this paper. We performed pre-processing steps to remove noise and intensity variances using N4ITK and standard intensity correction, respectively. We used Keras open-source library with Theano as backend for fast implementation of CNN model. In addition, we used BRATS 2015 MRI dataset to evaluate our proposed model. Furthermore, we have used SimpleITK open-source library in our proposed model to analyze images. Moreover, we have extracted random 2D patches for proposed 2D-CNN model for efficient brain segmentation. Extracting 2D patched instead of 3D due to less dimensional information present in 2D which helps us in reducing computational time. Dice Similarity Coefficient (DSC) is used as performance measure for the evaluation of the proposed method. Our method achieved DSC score of 0.77 for complete, 0.76 for core, 0.77 for enhanced tumor regions. However, these results are comparable with methods already implemented 2D CNN architecture.

Keywords: Deep learning, Convolutional Neural Networks, LGG, brain tumor segmentation

Procedia PDF Downloads 48
56 Understanding and Improving Neural Network Weight Initialization

Authors: Diego Aguirre, Olac Fuentes


In this paper, we present a taxonomy of weight initialization schemes used in deep learning. We survey the most representative techniques in each class and compare them in terms of overhead cost, convergence rate, and applicability. We also introduce a new weight initialization scheme. In this technique, we perform an initial feedforward pass through the network using an initialization mini-batch. Using statistics obtained from this pass, we initialize the weights of the network, so the following properties are met: 1) weight matrices are orthogonal; 2) ReLU layers produce a predetermined number of non-zero activations; 3) the output produced by each internal layer has a unit variance; 4) weights in the last layer are chosen to minimize the error in the initial mini-batch. We evaluate our method on three popular architectures, and a faster converge rates are achieved on the MNIST, CIFAR-10/100, and ImageNet datasets when compared to state-of-the-art initialization techniques.

Keywords: Deep learning, Supervised Learning, image classification, weight initialization

Procedia PDF Downloads 28