1,721,033 research outputs found
2nd International Conference on Robot Intelligence Technology and Applications
No full textWe are facing a new technological challenge on how to store and retrieve knowledge and manipulate intelligence for autonomous services by intelligent systems which should be capable of carrying out real world tasks autonomously. To address this issue, robot researchers have been developing intelligence technology (InT) for “robots that think” which is in the focus of this book. The book covers all aspects of intelligence from perception at sensor level and reasoning at cognitive level to behavior planning at execution level for each low level segment of the machine. It also presents the technologies for cognitive reasoning, social interaction with humans, behavior generation, ability to cooperate with other robots, ambience awareness, and an artificial genome that can be passed on to other robots. These technologies are to materialize cognitive intelligence, social intelligence, behavioral intelligence, collective intelligence, ambient intelligence and genetic intelligence. The book aims at serving researchers and practitioners with a timely dissemination of the recent progress on robot intelligence technology and its applications, based on a collection of papers presented at the at the 2nd International Conference on Robot Intelligence Technology and Applications (RiTA), held in Denver, USA, December 18-20, 2013
Exemplar-based Kernel Preserving Embedding
No full textWith the rapid increase of available data, it becomes computationally harder to extract useful information, specially in the case of high-dimensional data. Choosing a representative subset of the data can be useful to overcome this challenge as these representatives can be used by data analysts or presented to end users to give them a grasp of the data nature and structure.
In this dissertation, first an Exemplar-based approach for topic detection is proposed, in which detected topics are represented using a few selected tweets. Using exemplar tweets instead of a set of keywords allows for an easy interpretation of the meaning of the detected topics. The approach is then extended to detect topics that emerge in new epochs of data. Experimental evaluation on benchmark Twitter datasets shows that the proposed topic detection approach achieves the best term precision. It does this while maintaining good topic recall and running times compared to other approaches for topic detection. Moreover, the proposed emerging extension achieves higher topic recall with improved running times when compared to recent emerging topic detection approaches.
To overcome the challenge of high-dimensional data, several techniques, like PCA and NMF, were proposed to embed high-dimensional data into low-dimensional latent space. However, data represented in latent space is difficult for data analysts to understand and grasp the information encoded in it. In addition, these techniques do not take the relations between the data points into account. This motivated the development of other techniques like MDS, LLE and ISOMAP which preserve the relations between the data instances, but they still use latent features. In this dissertation, a new embedding technique is proposed to mitigate the previous problems by projecting the data to a space described by few points (i.e., the exemplars) which preserves the relations between the data points. The proposed method Exemplar-based Kernel Preserving (EBEK) embedding is shown theoretically to achieve the lowest reconstruction error of the kernel matrix. EBEK achieves a linear running time complexity in terms of the number of the samples. Using EBEK in the approximate nearest neighbor search task shows its ability to outperform related work by up to 60% in the recall while maintaining a good running time. In addition, empirical evaluation on clustering shows that EBEK achieves higher NMI than LLE and NMF by differences up to 40% and 15% respectively. It also achieves a comparable cluster quality to ISOMAP with a difference up to 3% in NMI and F-measure with a speedup up to 15×. In addition, our interpretability experiments show that EBEK’s selected basis are more understandable than the latent basis in images datasets
Versatile Deep Learning Forecasting Application with Metamorphic Quality Assurance
Get PDFAccurate estimates of fresh produce (FP) yields and prices are crucial for having fair bidding prices by retailers along with informed asking prices by farmers, leading to the best prices for customers. To have accurate estimates, the state-of-the-art deep learning (DL) models for forecasting FP yields and prices, including both station-based and satellite based models, are improved in this thesis by providing a new deep learning model structure.
The scope of this work covers forecasting a horizon of 5 weeks ahead for the fresh produce yields and prices. The proposed structure is built using an ensemble of Attention Deep Feedforward Neural Network with Gated Recurrent Units (ADGRU) and Deep Feedforward Neural Network with embedded GRU units (DFNNGRU); (DFNNGRU-ADGRU ENS).
The station-based version of the ensemble is trained and tested using as input the soil moisture and temperature parameters retrieved from land stations. This station-based ensemble model is found to outperform the literature model by 24% improvement in the AGM score for yield forecasting and 37.5% for price forecasting.
For the satellite-based model, the best satellite image preprocessing technique must be found to represent the images with less data for efficiency. Therefore, a preprocessing approach based on averaging is proposed and implemented then compared with the literature approach, which is based on histograms, where the proposed approach improves performance by 20%. The proposed Deep Feed Forward Neural Network with Embedded Gated Recurrent Units (DFNNGRU) ensembled with Attention Deep GRUs (ADGRU) is then tested against well-performing models of Stacked-AutoEncoder (SAE) ensembled with Convolution Neural Networks with Long-short term memory (CNNLSTM), where the proposed model is found to outperform the literature model by 12.5%.
In addition, interpolation techniques are used to estimate the missing VIs values due to the low frequency of capturing the satellite images by Landsat. A comparative analysis is conducted to choose the most effective technique, which is found to be Cubic Spline interpolation. The effect of adding the VIs as input parameters on the forecasting performance of the deep learning model is assessed and the most effective VIs are selected. One VI, which is the Normalized Difference Vegetation Index (NDVI), proves to be the most effective index in forecasting yield with an enhancement of 12.5% in AGM score.
A novel transfer learning (TL) framework is proposed for better generalizability. After finding the best DL forecasting model, a TL framework is proposed to enhance that model generalization to other FPs by using FP similarity, clustering, and TL techniques customized to fit the problem in hand. Furthermore, the similarity algorithms found in literature are improved by considering the time series features rather than the absolute values of their points. In addition, the FPs are clustered using a hierarchical clustering technique utilizing the complete linkage of a dendrogram to automate the process of finding the similarity thresholds and avoid setting them arbitrarily. Finally, the transfer learning is applied by freezing some layers of the proposed ensemble model and fine-tuning the rest leading to significant improvement in AGM compared to the best literature model.
Finally, a forecasting application is implemented to facilitate the use of the proposed models by the end users through a friendly interface. For testing the quality of the application deployed code and models, metamorphic testing is applied to assess the effectiveness of the machine learning models while machine learning is used to automatically detect the main metamorphic relations in the software code. The interactive role played by metamorphic testing and machine learning is investigated through the quality assurance of the forecasting application. The datasets used to train and test the deep learning forecasting models as well as the forecasting models are verified using metamorphic tests and the metamorphic relations in the generalization code are automatically detected using Support Vector Machine (SVM) models. Testing has revealed the unmatched requirements that are fixed to bring forward a valid application with sound data, effective models, and valid generalization code
AI-Driven Multi-Purpose Platform for Smart Healthcare Systems
No full textThe significant increase in the number of individuals with chronic ailments has dictated an urgent need for the development of an innovative model for healthcare systems. The design of smart healthcare platforms, a subject of recently growing interest, has become technologically feasible due to the emergence of modern technologies, including smartphones, wearable sensors, 5G communication networks, Internet of Things (IoT), cloud computing, and Artificial Intelligence (AI), particularly machine learning and deep learning. Together, these technologies enable expanded levels of data storage, computation, and secure communication for devices and servers, thus drastically increasing the degree of mobility, security, and available functionality. The thesis is focused on the development of AI-driven data fusion architecture in smart healthcare platform for a scalable, intelligent, and easily deployable remote monitoring system, aimed at providing cost-effective quality healthcare services in assisted living centers.
The contributions of the thesis are two-folds: we mainly focus on two major aspects of human behavior, including Human Activity Recognition (HAR) and emotion recognition to monitor the elderly in smart healthcare. In the first part, we propose two AI-enabled multimodal fusion approaches for HAR in ambient assisted living. We present a deep learning-based fusion approach for multimodal HAR that fuses the different modalities of data to obtain robust outcomes. Initially, Convolutional Neural Networks (CNNs) retrieve the high-level attributes from the image data, and the Convolutional Long Short Term Memory (ConvLSTM) is utilized to capture significant patterns from the multi-sensory data. Afterwards, the extracted features from the modalities are fused through self-attention mechanisms that enhance the relevant activity data and inhibit the superfluous and possibly confusing information by measuring their compatibility. Later, we propose a multi-level feature fusion technique for multimodal HAR using multi-head CNN with Convolution Block Attention Module (CBAM) to process the visual data and ConvLSTM for dealing with the time-sensitive multi-source sensor information. The architecture is developed to be able to analyze and retrieve channel and spatial dimension features through the use of three branches of CNN along with CBAM for visual information. The ConvLSTM network is designed to capture temporal features from the multiple sensors’ time-series data for efficient activity recognition.
The second part of this thesis focuses on multimodal emotion recognition in connected healthcare to monitor the patient’s health status. To achieve this goal, we introduce two feature fusion approaches through AI tools for emotion recognition. We propose a novel model-level fusion technique based on deep learning for enhanced emotion recognition from multimodal signals to monitor patients in connected healthcare. The representative visual features from the video signals are extracted through the Depthwise Separable Convolution Neural Network, and the optimized temporal attributes are derived from the multiple physiological data utilizing Bi-directional Long Short-Term Memory. A soft attention method fused the high multimodal features obtained from the two data modalities to retrieve the most significant features by focusing on emotionally salient parts of the features. We exploited two face detection methods, Histogram of Oriented Gradients and CNN-based face detector (ResNet-34), to observe the effects of facial features on emotion recognition. Lastly, we introduce a Multi-Stage Fusion Network (MSF-Net) for emotion recognition capable of extracting multimodal information and achieving significant performances. We propose utilizing the transformer-based structure to extract deep features from facial expressions. We exploited two visual descriptors, local binary pattern and Oriented FAST and Rotated BRIEF, to retrieve the computer vision-based features from the facial videos. A feature level fusion network integrates the extraction of features from these modules, directing the output into the triplet attention technique. This module employs a three-branch architecture to compute attention weights to capture cross-dimensional interactions efficiently. The temporal dependencies in physiological signals are modeled by a Bi-directional Gated Recurrent Unit (Bi-GRU) in forward and backward directions at each time step. Lastly, the output feature representations from the triplet attention module and the extracted high-level patterns from Bi-GRU are fused and fed into the classification module to recognize emotion.
Finally, we conduct a series of extensive experiments to demonstrate the performance against State-of-the-Art (SOTA) approaches. The findings from the experimental results reveal that the developed multimodal fusion networks surpass the existing SOTA methods in terms of multiple performance metrics. We deployed an IoT system to test the developed feature-fusion networks in real-world scalable smart healthcare application. The developed multimodal predictive analytics frameworks residing in the cloud and trained on large datasets are continually analyzed the nature of the ingested data, making any appropriate notifications to the patients themselves or the healthcare provider through a user-friendly human-machine interface
Noise Robust Keyword Spotting Using Deep Neural Networks For Embedded Platforms
No full textThe recent development of embedded platforms along with spectacular growth in communication networking technologies is driving the Internet of things to thrive. More complex tasks are now possible to operate in small devices such as speech recognition and keyword spotting which are in great demand. Traditional voice recognition approaches are already being used in several embedded applications, some are hybrid(cloud-based and embedded) while others are fully embedded. However, the environment surrounding the embedded devices is usually accompanied by noise. Conventional approaches to add noise robustness to speech recognition are effective but also costly in terms of memory consumption and hardware complexities which limit their use in embedded platforms. The purpose of this thesis is to increase the robustness of keyword spotting to more than one type of noise at once without increasing the memory footprint or the need for a denoiser while maintaining the recognition accuracy to an acceptable level. In this work, robustness in treated at the phoneme classification level as the phoneme based keyword spotting is the best technique for embedded keyword spotting.
Deep Neural Networks have been successfully deployed in many applications including noise robust speech recognition. In this work, we use mutil-condition utterances training of a Deep Neural Networks model to increase the keyword spotting noise robustness. This technique is also used for a Gaussian mixture model training. The two approaches are compared and the deep learning proved to not only outperform the Gaussian approach, but has also outperformed the use of a denoiser system. This results in a smaller, more accurate and noise robust model for phoneme recognition
Large-Scale Traffic Flow Prediction Using Deep Learning in the Context of Smart Mobility
Get PDFDesigning and developing a new generation of cities around the world (termed as smart cities) is fast becoming one of the ultimate solutions to overcome cities' problems such as population growth, pollution, energy crisis, and pressure demand on existing transportation infrastructure. One of the major aspects of a smart city is smart mobility. Smart mobility aims at improving transportation systems in several aspects: city logistics, info-mobility, and people-mobility. The emergence of the Internet of Car (IoC) phenomenon alongside with the development of Intelligent Transportation Systems (ITSs) opens some
opportunities in improving the tra c management systems and assisting the travelers and authorities in their decision-making process. However, this has given rise to the generation of huge amount of data originated from human-device and device-device interaction. This is an opportunity and a challenge, and smart mobility will not meet its full potential unless valuable insights are extracted from these big data.
Although the smart city environment and IoC allow for the generation and exchange of large amounts of data, there have not been yet well de ned and mature approaches for mining this wealth of information to bene t the drivers and traffic authorities. The main reason is most likely related to fundamental challenges in dealing with big data of various types and uncertain frequency coming from diverse sources. Mainly, the issues of types of data and uncertainty analysis in the predictions are indicated as the most challenging areas of study that have not been tackled yet. Important issues such as the nature of the data, i.e., stationary or non-stationary, and the prediction tasks, i.e., short-term or long-term, should also be taken into consideration. Based on this observation, a data-driven traffic flow prediction framework within the context of big data environment is proposed in this thesis. The main goal of this framework is to enhance the quality of traffic flow predictions, which can be used to assist travelers and traffic authorities in the decision-making process (whether for travel or management purposes).
The proposed framework is focused around four main aspects that tackle major data-driven traffic flow prediction problems: the fusion of hard data for traffic flow prediction; the fusion of soft data for traffic flow prediction; prediction of non-stationary traffic flow; and prediction of multi-step traffic flow. All these aspects are investigated and formulated as computational based tools/algorithms/approaches adequately tailored to the nature of the data at hand. The first tool tackles the inherent big data problems and deals with the uncertainty in the prediction. It relies on the ability of deep learning approaches in handling huge amounts of data generated by a large-scale and complex transportation system with limited prior knowledge. Furthermore, motivated by the close correlation between road traffic and weather conditions, a novel deep-learning-based approach that predicts traffic flow by fusing the traffic history and weather data is proposed.
The second tool fuses the streams of data (hard data) and event-based data (soft data) using Dempster Shafer Evidence Theory (DSET). One of the main features of the DSET is its ability to capture uncertainties in probabilities. Subsequently, an extension of DSET, namely Dempsters conditional rules for updating belief, is used to fuse traffic prediction beliefs coming from streams of data and event-based data sources. The third tool consists of a method to detect non-stationarities in the traffic flow and an algorithm to perform online adaptations of the tra c prediction model. The proposed detection approach is developed by monitoring the evolution of the spectral contents of the traffic flow. Furthermore, the approach is specfi cally developed to work in conjunction with state-of-the-art machine learning methods such as Deep Neural Network (DNN). By combining the power of frequency domain features and the known generalization capability and scalability of DNN in handling real-world data, it is expected that high prediction performances can be achieved.
The last tool is developed to improve multi-step traffic flow prediction in the recursive and multi-output settings. In the recursive setting, an algorithm that augments the information about the current time-step is proposed. This algorithm is called Conditional
Data as Demonstrator (C-DaD) and is an extension of an algorithm called Data as Demonstrator (DaD). Furthermore, in the multi-output setting, a novel approach of generating new history-future pairs of data that are aggregated with the original training data using Conditional Generative Adversarial Network (C-GAN) is developed.
To demonstrate the capabilities of the proposed approaches, a series of experiments using arti cial and real-world data are conducted. Each of the proposed approaches is compared with the state-of-the-art or currently existing approaches
Compression and Analysis of Pre-trained Language Model using Neural Slimming
Get PDFNeural networks are powerful solutions to help with decision making and solve complex
problems in recent years. In the domain of natural language processing, BERT and
its variants significantly outperform other network structures. It learns general linguistic
knowledge from a large corpus in the pre-training stage and utilizes that to solve downstream
tasks. However, the size of this kind of model is enormous and causes the issue
of overparameterization. This makes the model deployment on small edge devices less
scalable and flexible.
In this thesis, we study how to compress the BERT model in a structured pruning
manner. We proposed the neural slimming technique to assess the importance of each
neuron and designed the cost function and pruning strategy to remove neurons that make
zero or less contribution to the prediction. After getting fine-tuned on the downstream
tasks, the model can learn a more compact structure, and we name it SlimBERT.
We tested our method on 7 GLUE tasks and used only 10% original parameters to recover 94% of the original performance. It also reduced the run-time memory and increased the
inference speed at the same time. Compared to knowledge distillation methods and other
structured pruning methods, the proposed approach achieved better performance under
different metrics with the same compression ratio. Moreover, our method also improved
the interpretability of BERT. By analyzing neurons with a significant contribution, we can
observe that BERT utilizes different components and subnetworks according to different
tasks
Scalable Topic Detection Approaches fromTwitter Streams
No full textReal time topic detection in Twitter streams is an important task that helps discovering natural disasters in a real time from users’ posts and helps political parties and companies understand users’ opinions and needs. In 2014 the number of active users on Twitter is reported to be more than 288 million users who are posting around 500 million tweets daily. Therefore, detecting topics from Twitter streams in a real time becomes a challenging task that needs scalable and efficient techniques to handle this large amount of data. In this work, we scale an Exemplar-based technique that detects topics from Twitter streams, where each of the detected topics is represented by one tweet (i.e, exemplar). Using exemplar tweets to represent the detected topics, makes these topics easier to interpret as opposed to representing them by uncorrelated terms as in other topic detection algorithms. The approach is implemented using Apache Giraph and is being extended here to efficiently support sliding windows. Experimental results on four datasets show that the optimized Giraph implementation achieves a speedup of up to nineteen times over the
native implementation, while maintaining good quality of the detected topics. In addition, Giraph Exemplar-based approach achieves the best topic recall and term precision against K-means, Latent Dirichlet Allocation (LDA), Non-negative matrix factorization (NMF) and Latent Semantic Analysis (LSA), while maintaining a good term recall and running time. The approach is also deployed for detecting topics from real-time Twitter streams and its scalability is demonstrated.
Moreover, another clustering technique called Local Variance-based Clustering (LVC) is proposed in this thesis for detecting topics from Twitter streams. Local Variance-based Clustering (LVC) defines the data points densities based on their similarities. The proposed local variance measure is calculated based on the variance of the data points similarity histogram and is shown to well distinguish between core, border, connecting and outliers points. Experimental results show that LVC outperforms spectral clustering and affinity propagation in clustering quality using control charts, Ecoli and images datasets, while maintaining a good running time. In addition, results show that LVC can detect topics from Twitter with higher topic recall by 15% and higher
term precision by 3% over DBSCAN
Continual Learning and Forgetting in Deep Learning Models
Get PDFContinual learning is a framework of learning in which we aim to move beyond the limitations of standard isolated optimization of deep learning models toward a more intelligent setting, where models or agents are able to accumulate skills and knowledge, across diverse tasks and over extended periods of time, much like humans do. Like much of neural networks research, interest in continual learning has ebbed and flowed over the decades, and ultimately saw a sharp increase over the past few years, buoyed by the successes of deep learning thus far.
One obstacle that has dominated continual learning research over the years is the so-called catastrophic forgetting phenomenon, which refers to the tendency of neural networks to "forget" older skills and knowledge as soon as they are subsequently optimized for additional tasks. Researchers have proposed various approaches to counter forgetting in neural networks. In this dissertation, we review some of those approaches and build upon them, and address other aspects of the continual learning problem.
We make the following four contributions.
First, we address the critical role of importance estimation in fixed-capacity models, where the aim is to find a balance between countering forgetting and preserving a model's capacity to learn additional tasks. We propose a novel unit importance estimation approach, with a small memory and computational footprint. The proposed approach builds on recent work that showed that the average of a unit's activation values is a good indicator of its importance, and extends it by taking into consideration the separation between class-conditional distributions of activation values.
Second, we observe that most methods that aim to prevent forgetting by explicitly penalizing changes to parameters can be seen as post hoc remedies that ultimately lead to inefficient use of model capacity. We argue that taking into account the continual learning objective requires a modification to the optimization approach from the start rather than only after learning. In particular, we argue that key to the effective use of a model's capacity in the continual learning setting is to drive the optimization process toward learning more general, reusable, and thus durable representations that are less susceptible to forgetting. To that end, we explore the use of supervised and unsupervised auxiliary tasks as regularization, not against forgetting, but against learning representations that narrowly target any single classification task. We show that the approach is successful at mitigating forgetting, even though it does not explicitly penalize forgetting.
Third, we explore the effect of inter-task similarity in sequences of image classification tasks on the overall performance of continual learning models. We show that certain models are adversely affected when the learned tasks are dissimilar. Moreover, we show that, in those cases, a small replay memory, even 1% the size of the training data, is enough to significantly improve performance.
Fourth and lastly, we explore the performance of continual learning models in the so-called multi-head and single-head settings and approaches to narrow the gap between the two settings. We show that unlabelled auxiliary data, not sampled from any task in the learning sequence, can be used to improve performance in the single-head setting.
We provide extensive empirical evaluation of the proposed approaches and compare their performance against recent continual learning methods in the literature
- …
