Dissertations / Theses on the topic 'LSTM Temporel'

Ce travail de thèse s'inscrit dans la catégorie de la recherche en informatique de santé, en particulier l'analyse et la prédiction des parcours patients, qui sont les séquences des actes médicaux consommés par les patients au fil du temps. Notre objectif est de proposer une approche innovante pour l'exploitation des données de parcours de soins afin de réaliser non seulement une classification binaire, mais aussi multi-label. Nous concevons également une nouvelle approche de vectorisation et représentation sémantique exclusivement pour le domaine médical français, qui permettra d'exploiter un autre aspect des parcours patients afin d'améliorer la performance prédictive de notre approche proposée. Notre recherche s'inscrit dans le cadre des travaux de CEGEDIM ASSURANCES, une business unit du groupe CEGEDIM qui fournit des logiciels et des services pour les secteurs de l'assurance maladie complémentaire et de la gestion des risques en France. En analysant le parcours de soins et en utilisant l'approche que nous proposons, nous pouvons extraire des informations précieuses et identifier des patterns dans les parcours médicaux des patients afin de prédire des événements médicaux potentiels ou la consommation médicale à venir. Cela permettra aux assureurs de prévoir les futures demandes de soins de santé et donc de négocier de meilleurs tarifs avec les prestataires de soins de santé, ce qui permettra une planification financière précise, des modèles de tarification équitables et une réduction des coûts. En outre, ça permettra aux assureurs privés de concevoir des plans de santé personnalisés qui répondent aux besoins spécifiques des patients, en veillant à ce qu'ils reçoivent les soins adéquats au bon moment afin de prévenir la progression de la maladie. Enfin, l'offre de programmes de soins préventifs et de produits et services de santé personnalisés renforce les relations avec les clients, améliore leur satisfaction et réduit l'attrition. Dans ce travail, nous visons à développer une approche permettant d'analyser les parcours patients et de prédire les événements médicaux ou les traitements à venir, sur la base d'un large portefeuille de remboursements. Pour atteindre cet objectif, nous proposons tout d'abord un nouveau modèle basé sur les LSTM qui tient compte de la notion temporelle et qui permet de réaliser de la classification binaire et multi-label. Le modèle proposé est ensuite étendu par un autre aspect des parcours de soins, à savoir des informations supplémentaires provenant d'un clustering flou du même portefeuille. Nous démontrons que l'approche proposée est plus performante que les méthodes traditionnelles et d'apprentissage profond dans la prédiction médicale binaire et multi-label. Par la suite, nous améliorons la performance prédictive de l'approche proposée en exploitant un aspect supplémentaire des parcours patients, qui consiste en une description textuelle détaillée des traitements médicaux consommés. Ceci est réalisé grâce à la conception de F-BERTMed, une nouvelle approche de vectorisation et de représentation sémantique de phrases pour le domaine médical français. Celle-ci présente des avantages significatifs par rapport aux méthodes de l'état de l'art du traitement automatique du langage naturel (TAL). F-BERTMed est basé sur FlauBERT, dont le pré-entraînement utilisant la tâche MLM (Modélisation Masqué du Langage) a été étendu sur des textes médicaux français avant d'être fine-tuné sur les tâches NLI (Inférence du Langage Naturel) et STS (Similarité Sémantique Textuelle). Nous démontrons enfin que l'utilisation de F-BERTMed pour générer une nouvelle représentation des parcours patients améliore les performances prédictives de notre modèle proposé pour les tâches de classification binaire et multi-label
This thesis work falls into the category of healthcare informatics research, specifically the analysis and prediction of patients’ care pathways, which are the sequences of medical services consumed by patients over time. Our aim is to propose an innovative approach for the exploitation of patient care trajectory data in order to achieve not only binary, but also multi-label classification. We also design a new sentence embedding framework exclusively for the french medical domain, which will harness another view of the patients’ care pathways in order to enhance the predictive performance of our proposed approach. Our research is part of the work of CEGEDIM ASSURANCES, a business unit of the CEGEDIM Group that provides software and services for the french supplementary healthcare insurance and risk management sectors. By analyzing the patient care pathway and leveraging our proposed approach, we can extract valuable insights and identify patterns within the patients’ medical journeys in order to predict potential medical events or upcoming medical consumption. This will allow insurers to forecast future healthcare claims and therefore negotiate better rates with healthcare providers, allowing for accurate financial planning, fair pricing models and cost reductions. Furthermore, it enables private healthcare insurers to design personalized health plans that meet the specific needs of the patients, ensuring they receive the right care at the right time to prevent disease progression. Ultimately, offering preventive care programs and customized health products and services enhances client relationship, improving their satisfaction and reducing churn. In this work, we aim to develop an approach to analyze patient care pathways and predict medical events or upcoming treatments, based on a large portfolio of reimbursed medical records. To achieve this goal, we first propose a new time-aware long-short term memory based framework that can achieve both binary and multi-label classification. The proposed framework is then extended with another aspect of the patient healthcare trajectories, namely additional information from a fuzzy clustering of the same portfolio. We show that our proposed approach outperforms traditional and deep learning methods in medical binary and multi-label prediction. Subsequently, we enhance the predictive performance of our proposed approach by exploiting a supplementary view of the patient care pathways that consists of a detailed textual description of the consumed medical treatments. This is achieved through the design of F-BERTMed, a new sentence embedding framework for the french medical domain that presents significant advantages over the natural language processing (NLP) state-of-the-art methods. F-BERTMed is based on FlauBERT, whose pre-training using MLM (Masked Language Modeling) was extended on french medical texts before being fine-tuned on NLI (Natural Language Inference) and STS (Semantic Textual Similarity) tasks. We finally show that using F-BERTMed to generate a new representation of the patient care pathways enhances the performance of our proposed medical predictive framework on both binary and multi-label classification tasks

We explore the use of Long short-term memory (LSTM) for anomaly detection in temporal data. Due to the challenges in obtaining labeled anomaly datasets, an unsupervised approach is employed. We train recurrent neural networks (RNNs) with LSTM units to learn the normal time series patterns and predict future values. The resulting prediction errors are modeled to give anomaly scores. We investigate different ways of maintaining LSTM state, and the effect of using a fixed number of time steps on LSTM prediction and detection performance. LSTMs are also compared to feed-forward neural networks with fixed size time windows over inputs. Our experiments, with three real-world datasets, show that while LSTM RNNs are suitable for general purpose time series modeling and anomaly detection, maintaining LSTM state is crucial for getting desired results. Moreover, LSTMs may not be required at all for simple time series.
Vi undersöker Long short-term memory (LSTM) för avvikelsedetektion i tidsseriedata. På grund av svårigheterna i att hitta data med etiketter så har ett oövervakat an-greppssätt använts. Vi tränar rekursiva neuronnät (RNN) med LSTM-noder för att lära modellen det normala tidsseriemönstret och prediktera framtida värden. Vi undersö-ker olika sätt av att behålla LSTM-tillståndet och effekter av att använda ett konstant antal tidssteg på LSTM-prediktionen och avvikelsedetektionsprestandan. LSTM är också jämförda med vanliga neuronnät med fasta tidsfönster över indata. Våra experiment med tre verkliga datasetvisar att även om LSTM RNN är tillämpbara för generell tidsseriemodellering och avvikelsedetektion så är det avgörande att behålla LSTM-tillståndet för att få de önskaderesultaten. Dessutom är det inte nödvändigt att använda LSTM för enkla tidsserier.

The low amount solved residential burglary crimes calls for new and innovative methods in the prevention and investigation of the cases. There were 22 600 reported residential burglaries in Sweden 2017 but only four to five percent of these will ever be solved. There are many initiatives in both Sweden and abroad for decreasing the amount of occurring residential burglaries and one of the areas that are being tested is the use of prediction methods for more efficient preventive actions. This thesis is an investigation of a potential method of prediction by using neural networks to identify areas that have a higher risk of burglaries on a daily basis. The model use reported burglaries to learn patterns in both space and time. The rationale for the existence of patterns is based on near repeat theories in criminology which states that after a burglary both the burgled victim and an area around that victim has an increased risk of additional burglaries. The work has been conducted in cooperation with the Swedish Police authority. The machine learning is implemented with convolutional long short-term memory (LSTM) neural networks with max pooling in three dimensions that learn from ten years of residential burglary data (2007-2016) in a study area in Stockholm, Sweden. The model's accuracy is measured by performing predictions of burglaries during 2017 on a daily basis. It classifies cells in a 36x36 grid with 600 meter square grid cells as areas with elevated risk or not. By classifying 4% of all grid cells during the year as risk areas, 43% of all burglaries are correctly predicted. The performance of the model could potentially be improved by further configuration of the parameters of the neural network, along with a use of more data with factors that are correlated to burglaries, for instance weather. Consequently, further work in these areas could increase the accuracy. The conclusion is that neural networks or machine learning in general could be a powerful and innovative tool for the Swedish Police authority to predict and moreover prevent certain crime. This thesis serves as a first prototype of how such a system could be implemented and used.

Due to automation, the world is changing at a rapid pace. Autonomous agents have become more common over the last several years and, as a result, have created a need for improved software to back them up. The most important aspect of this greater software is path prediction, as robots need to be able to decide where to move in the future. In order to accomplish this, a robot must know how to avoid humans, putting frame prediction at the core of many modern day solutions. A popular way to solve this complex problem of frame prediction is Auto Encoder LSTMs. Though there are many implementations of this, at its core, it is a neural network comprised of a series of time sensitive processing blocks that shrink and then grow the data’s dimensions to make a prediction. The idea of using Auto Encoder styled networks to do frame prediction has also been adapted by others to make Temporal Encoders. These neural networks work much like traditional Auto Encoders, in which the data is reduced then expanded back up. These networks attempt to tease out a series of frames, including a predictive frame of the future. The problem with many of these networks is that they take an immense amount of computation power, and time to get them performing at an acceptable level. This thesis presents possible ways of pre-processing input frames to these networks in order to gain performance, in the best case seeing a 360x improvement in accuracy compared to the original models. This thesis also extends the work done with Temporal Encoders to create more precise prediction models, which showed consistent improvements of at least 50% for some metrics. All of the generated models were compared using a simulated data set collected from recordings of ground level viewpoints from Cities: Skylines. These predicted frames were then analyzed using a common perceptual distance metric, that is, Minkowski distance, as well as a custom metric that tracked distinct areas in frames. All of the following was run on a constrained system in order to see the effects of the changes as they pertain to systems with limited hardware access.

Recently, the availability of high quality and high resolution spatio-temporal data has increased for many sports. This enabled deep analysis of player behaviour and game strategy. This thesis investigates the assumption that game strategy is latent information in tracking data from soccer games and the possibility of modelling player behaviour with deep imitation learning. A possible application would be to perform counterfactual analysis, and switch an observed player in a real sequence, with a simulated player to asses alternative scenarios. An imitation learning application is implemented using recurrent neural networks. It is shown that the application is able to learn individual player behaviour and perform rollouts on previously unseen sequences.

En médecine prédictive personnalisée, modéliser avec précision la maladie et les processus de soins d'un patient est crucial en raison des dépendances temporelles à long terme inhérentes. Cependant, les dossiers de santé électroniques (DSE) se composent souvent de données épisodiques et irrégulières, issues des admissions hospitalières sporadiques, créant des schémas uniques pour chaque séjour hospitalier.Par conséquent, la construction d'un modèle prédictif personnalisé nécessite une considération attentive de ces facteurs pour capturer avec précision le parcours de santé du patient et aider à la prise de décision clinique.LSTM sont efficaces pour traiter les données séquentielles comme les DSE, mais ils présentent deux limitations majeures : l'incapacité à interpréter les résultats des prédictions et à prendre en compte des intervalles de temps irréguliers entre les événements consécutifs. Pour surmonter ces limitations, nous introduisons de nouveaux réseaux neuronaux à mémoire dynamique profonde appelés Multi-Way Adaptive et Adaptive Multi-Way Interpretable Time-Aware LSTM (MWTA-LSTM etAMITA), conçus pour les données séquentielles collectées de manière irrégulière.L'objectif principal des deux modèles est de tirer parti des dossiers médicaux pour mémoriser les trajectoires de maladie et les processus de soins, estimer les états de maladie actuels et prédire les risques futurs, offrant ainsi un haut niveau de précision et de pouvoir prédictif
In personalized predictive medicine, accurately modeling a patient's illness and care processes is crucial due to the inherent long-term temporal dependencies. However, Electronic Health Records (EHRs) often consist of episodic and irregularly timed data, stemming from sporadic hospital admissions, which create unique patterns for each hospital stay. Consequently, constructing a personalized predictive model necessitates careful consideration of these factors to accurately capture the patient's health journey and assist in clinical decision-making. LSTM networks are effective for handling sequential data like EHRs, but they face two significant limitations: the inability to interpret prediction results and to take into account irregular time intervals between consecutive events. To address limitations, we introduce novel deep dynamic memory neural networks called Multi-Way Adaptive and Adaptive Multi-Way Interpretable Time-Aware LSTM (MWTA-LSTM and AMITA) designed for irregularly collected sequential data. The primary objective of both models is to leverage medical records to memorize illness trajectories and care processes, estimate current illness states, and predict future risks, thereby providing a high level of precision and predictive power

Polyphonic or multiple music instrument detection is a difficult problem compared to detecting single or solo instruments in an audio recording. As music is time series data it be can modelled using sequence learning methods within deep learning. Recently, temporal convolutional networks (TCN) have shown to outperform conventional recurrent neural networks (RNN) on various sequence modelling tasks. Though there have been significant improvements in deep learning methods, data scarcity becomes a problem in training large scale models. Weakly labelled data is an alternative where a clip is annotated for presence or absence of instruments without specifying the times at which an instrument is sounding. This study investigates how TCN model compares to a Long Short-Term Memory (LSTM) model while trained on weakly labelled dataset. The results showed successful training of both models along with generalisation on a separate dataset. The comparison showed that TCN performed better than LSTM, but only marginally. Therefore, from the experiments carried out it could not be explicitly concluded if TCN is convincingly a better choice over LSTM in the context of instrument detection, but definitely a strong alternative.
Polyfonisk eller multipel musikinstrumentdetektering är ett svårt problem jämfört med att detektera enstaka eller soloinstrument i en ljudinspelning. Eftersom musik är tidsseriedata kan den modelleras med hjälp av sekvensinlärningsmetoder inom djup inlärning. Nyligen har ’Temporal Convolutional Network’ (TCN) visat sig överträffa konventionella ’Recurrent Neural Network’ (RNN) på flertalet sekvensmodelleringsuppgifter. Även om det har skett betydande förbättringar i metoder för djup inlärning, blir dataknapphet ett problem vid utbildning av storskaliga modeller. Svagt märkta data är ett alternativ där ett klipp kommenteras för närvaro av frånvaro av instrument utan att ange de tidpunkter då ett instrument låter. Denna studie undersöker hur TCN-modellen jämförs med en ’Long Short-Term Memory’ (LSTM) -modell medan den tränas i svagt märkta datasätt. Resultaten visade framgångsrik utbildning av båda modellerna tillsammans med generalisering i en separat datasats. Jämförelsen visade att TCN presterade bättre än LSTM, men endast marginellt. Därför kan man från de genomförda experimenten inte uttryckligen dra slutsatsen om TCN övertygande är ett bättre val jämfört med LSTM i samband med instrumentdetektering, men definitivt ett starkt alternativ.

Visual Object Tracking is the task of tracking an object within a video. Broadly, most tracking algorithms can be classified into neural network based, correlation filter based, and hybrid. This thesis investigates various methods to improve tracking using correlation filters. The thesis contributes four novel trackers. The first tracker uses an appearance model pool to avoid faulty filter updates. Next, the appearance feature channel weights are learned using the graph-based similarity followed by modelling sparse spatio-temporal variations. At last, non-linearity of the appearance features is captured. The thesis also presents extensive evaluation of the proposed trackers on standard datasets.

L'utilizzo dell'Intelligenza Artificiale in ambito industriale sta prendendo piede negli ultimi anni e il caso studiato in questa tesi ne è la prova. Lo sviluppo della tecnologia ha reso disponibile sempre più potenza computazionale a minor prezzo, rendendo possibile l'utilizzo delle Reti Neurali Profonde, studiate fin dagli anni ottanta, in un modo che fino a non molti anni fa era economicamente insostenibile. Si andrà a vedere il caso concreto della realizzazione di un sistema che esegue previsioni in tempo reale su telemetrie di un impianto per la gestione delle acque, con lo scopo di assistere gli operatori nelle decisioni critiche da prendere in situazioni che potrebbero portare a un'emergenza. Sono state utilizzate tecniche allo stato dell'arte del Deep Learning per la realizzazione della rete previsionale, soluzioni di Big Data e Cloud Computing per la raffinazione dei dati grezzi e rendere possibile il training della rete neurale. Sono state studiate le basi teoriche richieste per realizzare un sistema in streaming, è stata poi progettata e realizzata una architettura apposita dedicata alla trasformazione in tempo reale dei dati per poter realizzare previsioni aggiornate.

This thesis examines two different approaches to parsing event-based vehicular signal data to produce input to a neural network prediction model: event parsing, where the data is kept unevenly spaced over the temporal domain, and slice parsing, where the data is made to be evenly spaced over the temporal domain instead. The dataset used as a basis for these experiments consists of a number of vehicular signal logs taken at Scania AB. Comparisons between the parsing methods have been made by first training long short-term memory (LSTM) recurrent neural networks (RNN) on each of the parsed datasets and then measuring the output error and resource costs of each such model after having validated them on a number of shared validation sets. The results from these tests clearly show that slice parsing compares favourably to event parsing.
Denna avhandling jämför två olika tillvägagångssätt vad gäller parsningen av händelsebaserad signaldata från fordon för att producera indata till en förutsägelsemodell i form av ett neuronnät, nämligen händelseparsning, där datan förblir ojämnt fördelad över tidsdomänen, och skivparsning, där datan är omgjord till att istället vara jämnt fördelad över tidsdomänen. Det dataset som används för dessa experiment är ett antal signalloggar från fordon som kommer från Scania. Jämförelser mellan parsningsmetoderna gjordes genom att först träna ett lång korttidsminne (LSTM) återkommande neuronnät (RNN) på vardera av de skapade dataseten för att sedan mäta utmatningsfelet och resurskostnader för varje modell efter att de validerats på en delad uppsättning av valideringsdata. Resultaten från dessa tester visar tydligt på att skivparsning står sig väl mot händelseparsning.

碩士
國立清華大學
資訊工程學系所
106
``Is there any pattern in location-based, mobile check-in activities?'' ``If yes, is it possible to accurately predict the intention of a user's next check-in, given his/her check-in history?'' To answer these questions, we crawl and analyze probably the largest mobile check-in datasets, containing 20 million check-in activities from 0.4 million users. We provide two observations---`` work-n-relax'' and ``diurnal-n-nocturnal''---showing that the intentions of users' check-ins are strongly associated with time. Furthermore, the category of each check-in venue, which reveals users' intentions, has structure and forms taxonomy. In this paper, we propose Nested LSTM that takes both (a) check-in time and (b) taxonomy structure of venues from check-in sequences into consideration, providing accurate predictions on the category of a user's next check-in location. Nested LSTM also projects each category into an embedding space, providing a new representation with stronger semantic meanings. Experimental results are poised to demonstrate the effectiveness of the proposed Nested LSTM: (a) Nested LSTM improves Accuracy@5 by 4.22% on average, and (b) Nested LSTM learns a better taxonomy embedding for clustering categories, which improves Silhouette Coefficient by 1.5X. Both results (a)(b) are compared with LSTM-based, state-of-the-art approaches.

碩士
國立交通大學
資訊科學與工程研究所
107
The EEG signal is a medium to realize a brain-computer interface (BCI) system which help motor-disabled patients to communicate with the outside world by external devices. The problems associated with this task include recordings with a poor signal-to-noise ratio and contamination from external body movements, such as, muscle activity, blinking, and head movement. Considerable variability between subjects and recording sessions compounds the difficulty of this task, particularly when seeking to train a model using trials obtained from all of the subjects. Recently, there are works demonstrating the postive outcome using CNN in task of motor-imagery classification. This paper outlines two novel neural network architecture for the classification of motor imagery EEG recordings using deep learning techniques. One of proposed methods comprises an attention mechanism, the another model is CNN equipped with LSTM. The attenion mechanism in the former model calculating the importance of each electrode; the LSTM in the latter model used for finding the temporal information within features. Compared to the results obtained using a variety of state-of-the-art deep learning techniques, the proposed scheme represents a considerable advancement in classification accuracy when applied to the BCI Competition IVdataset IIa, reaching accuracy 85.2%. Besides, when the proposed models were applied to motor-imagery EEG data collected in this work, the models yielded better results compared to pure CNN model by 9.2%. Asides from comparing the accuracy to effectiveness of the proposed models, we also determine that the attention mechanism mentioned above performs the same process as CSP and common temporal pattern (CTP), wherein inputs from all classes are projected onto a similar coordinate system considered the optimal space for classification. Moreover, through power-feature corrlation maps, visualzation of LSTM, and representation erasure determined by RL, we rationalize the semantic meanings behind operations of CNN as well as LSTM and, eventually, illustrate out two decisive factors of temporal features affecting the capability of LSTM in sequence modeling: (i) critical time range for classification and (ii) correct frequency range for event-related potential (ERP) which induces the activation of the features. These two factors could be indications for designing models consisted of CNN and RNN for processing other types of bio-signal which are also closely in relationship with ERP.

Dissertation submitted in partial fulfillment of the requirements for the Degree of Master of Science in Geospatial Technologies.
The semiarid and arid areas of the eastern part of Ethiopia have suffered a series of droughts and famines in the years 1999/2000, 2003/2004, 2007 and 2011. Absence/decline of rainfall in two of the rainy seasons locally called Dihra and Gu as being the major fact behind drought. Besides, lack of appropriate monitoring techniques aggravate the situation of drought in the study area. In a region where the numbers of meteorological stations are not sufficient enough to monitor the onset and extent of drought, remotely sensed data presents fast and economical way of information as the ground condition reflects the overall condition of rainfall and soil moisture. In this study, the drought monitoring approach is developed using Terra-MODIS Normalized Difference Vegetation index (NDVI) and Land surface Temperature (LST) level-3 products. The approach integrates the land surface reflectance and thermal properties as well as the NDVI changes to identify the extent and pattern of the past drought years. From the NDVI versus LST scatter plot, we extract Vegetation Temperature condition index (VTCI) to map the variability and trend of the drought years. The year 2003 was found to be the driest year (more than 90% of the region affected by drought) and the season that showed increasing intensity of drought being Dihra. The correlation (r > 0.7) between rainfall and VTCI across the major meteorological stations suggested that the index could be used as good indicator of drought as rainfall does. The overall trend of drought condition for selected drought years suggested that eastern and southern regions will experience more severe drought in the coming year. Moreover, VTCI value for October from 2000-2011 showed similar increase intensity of drought condition. In addition, it was observed that sparse vegetation and shrub land are highly variable and bare soil region is consistently dry. Wetter regions were found in the area where the elevation is above 1500m above sea level.