Academic literature on the topic 'LSTM Neural networks'

LSTM neural networks have been used for speech recognition, image recognition and other artificial intelligence applications for many years. Most applications perform the LSTM algorithm and the required calculations on cloud computers. Off-line solutions include the use of FPGAs and GPUs but the most promising solutions include ASIC accelerators designed for this purpose only. This report presents an ASIC design capable of performing the multiple iterations of the LSTM algorithm on a unidirectional and without peepholes neural network architecture. The proposed design provides arithmetic level parallelism options as blocks are instantiated based on parameters. The internal structure of the design implements pipelined, parallel or serial solutions depending on which is optimal in every case. The implications concerning these decisions are discussed in detail in the report. The design process is described in detail and the evaluation of the design is also presented to measure accuracy and error of the design output.This thesis work resulted in a complete synthesizable ASIC design implementing an LSTM layer, a Fully Connected layer and a Softmax layer which can perform classification of data based on trained weight matrices and bias vectors. The design primarily uses 16-bit fixed point format with 5 integer and 11 fractional bits but increased precision representations are used in some blocks to reduce error output. Additionally, a verification environment has also been designed and is capable of performing simulations, evaluating the design output by comparing it with results produced from performing the same operations with 64-bit floating point precision on a SystemVerilog testbench and measuring the encountered error. The results concerning the accuracy and the design output error margin are presented in this thesis report. The design went through Logic and Physical synthesis and successfully resulted in a functional netlist for every tested configuration. Timing, area and power measurements on the generated netlists of various configurations of the design show consistency and are reported in this report.<br>LSTM neurala nätverk har använts för taligenkänning, bildigenkänning och andra artificiella intelligensapplikationer i många år. De flesta applikationer utför LSTM-algoritmen och de nödvändiga beräkningarna i digitala moln. Offline lösningar inkluderar användningen av FPGA och GPU men de mest lovande lösningarna inkluderar ASIC-acceleratorer utformade för endast dettaändamål. Denna rapport presenterar en ASIC-design som kan utföra multipla iterationer av LSTM-algoritmen på en enkelriktad neural nätverksarkitetur utan peepholes. Den föreslagna designed ger aritmetrisk nivå-parallellismalternativ som block som är instansierat baserat på parametrar. Designens inre konstruktion implementerar pipelinerade, parallella, eller seriella lösningar beroende på vilket anternativ som är optimalt till alla fall. Konsekvenserna för dessa beslut diskuteras i detalj i rapporten. Designprocessen beskrivs i detalj och utvärderingen av designen presenteras också för att mäta noggrannheten och felmarginal i designutgången. Resultatet av arbetet från denna rapport är en fullständig syntetiserbar ASIC design som har implementerat ett LSTM-lager, ett fullständigt anslutet lager och ett Softmax-lager som kan utföra klassificering av data baserat på tränade viktmatriser och biasvektorer. Designen använder huvudsakligen 16bitars fast flytpunktsformat med 5 heltal och 11 fraktions bitar men ökade precisionsrepresentationer används i vissa block för att minska felmarginal. Till detta har även en verifieringsmiljö utformats som kan utföra simuleringar, utvärdera designresultatet genom att jämföra det med resultatet som produceras från att utföra samma operationer med 64-bitars flytpunktsprecision på en SystemVerilog testbänk och mäta uppstådda felmarginal. Resultaten avseende noggrannheten och designutgångens felmarginal presenteras i denna rapport.Designen gick genom Logisk och Fysisk syntes och framgångsrikt resulterade i en funktionell nätlista för varje testad konfiguration. Timing, area och effektmätningar på den genererade nätlistorna av olika konfigurationer av designen visar konsistens och rapporteras i denna rapport.

Long short-term memory (LSTM) neural networks have been proven to be effective for time series prediction, even in some instances where the data is non-stationary. This lead us to examine their predictive ability of stock market returns, as the development of stock prices and returns tend to be a non-stationary time series. We used daily stock trading data to let an LSTM train models at predicting daily returns for 60 stocks from the OMX30 and Nasdaq-100 indices. Subsequently, we measured their accuracy, precision, and recall. The mean accuracy was 49.75 percent, meaning that the observed accuracy was close to the accuracy one would observe by randomly selecting a prediction for each day and lower than the accuracy achieved by blindly predicting all days to be positive. Finally, we concluded that further improvements need to be made for models trained by LSTMs to have any notable predictive ability in the area of stock returns.

Research in asset pricing has, until recently, side-stepped the high dimensionality problem by focusing on low-dimensional models. Work on cross-sectional stock return prediction, for example, has focused on regressions with a small number of characteristics. Given the background of an enormously large number of variables that could potentially be relevant for predicting returns, focusing on such a small number of factors effectively means that the researchers are imposing a very high degree of sparsity on these models. This research studies the use of the recurrent neural network (RNN) method to deal with the “curse of dimensionality” challenge in the cross-section of stock returns. The purpose is to predict the daily stock returns. Compared with the traditional method of returns, namely the CAPM model, the focus will be on using the LSTM model to do the prediction. LSTM is very powerful in sequence prediction problems because they’re able to store past information. Thus, we compare the forecast of returns from the LSTM model with the traditional CAPM model. The comparison will be made using the out-of-sample R2 along with the Sharpe Ratio and Sortino Ratio. Finally, we conclude with the further improvements that need to be made for models trained by LSTMs to have any notable predictive ability in the area of stock returns.

The harbour porpoise is a toothed whale whose presence is threatened in Scandinavia. Onestep towards preserving the species in critical areas is to study and observe the harbourporpoise population growth or decline in these areas. Today this is done by using underwateraudio recorders, so called hydrophones, and manual analyzing tools. This report describes amethod that modernizes the process of harbour porpoise detection with machine learning. Thedetection method is based on data collected by the hydrophone AQUAclick 100. The data isprocessed and classified automatically with a stacked long short-term memory recurrent neuralnetwork designed specifically for this purpose.<br>Vanlig tumlare är en tandval vars närvaro i Skandinavien är hotad. Ett steg mot att kunnabevara arten i utsatta områden är att studera och observera tumlarbeståndets tillväxt ellertillbakagång i dessa områden. Detta görs idag med hjälp av ljudinspelare för undervattensbruk,så kallade hydrofoner, samt manuella analysverktyg. Den här rapporten beskriver enmetod som moderniserar processen för detektering av vanlig tumlare genom maskininlärning.Detekteringen är baserad på insamlad data från hydrofonen AQUAclick 100. Bearbetning ochklassificering av data har automatiserats genom att använda ett staplat återkopplande neuraltnätverk med långt korttidsminne utarbetat specifikt för detta ändamål.

Interpreting time varying phenomena is a key challenge in the capital markets. Time series analysis using autoregressive methods has been carried out over the last couple of decades, often with reassuring results. However, such methods sometimes fail to explain trends and cyclical fluctuations, which may be characterized by long-range dependencies or even dependencies between the input features. The purpose of this thesis is to investigate whether recurrent neural networks with LSTM-cells can be used to capture these dependencies, and ultimately be used as a complement for index trading decisions. Experiments are made on different setups of the S&amp;P-500 stock index, and two distinct models are built, each one being an improvement of the previous model. The first model is a multivariate regression model, and the second model is a multivariate binary classifier. The output of each model is used to reason about the future behavior of the index. The experiment shows for the configuration provided that LSTM RNNs are unsuitable for predicting exact values of daily returns, but gives satisfactory results when used to predict the direction of the movement.<br>Att förstå och kunna förutsäga hur index varierar med tiden och andra parametrar är ett viktigt problem inom kapitalmarknader. Tidsserieanalys med autoregressiva metoder har funnits sedan årtionden tillbaka, och har oftast gett goda resultat. Dessa metoder saknar dock möjligheten att förklara trender och cykliska variationer i tidsserien, något som kan karaktäriseras av tidsvarierande samband, men även samband mellan parametrar som indexet beror utav. Syftet med denna studie är att undersöka om recurrent neural networks (RNN) med long short-term memory-celler (LSTM) kan användas för att fånga dessa samband, för att slutligen användas som en modell för att komplettera indexhandel. Experimenten är gjorda mot en modifierad S&amp;P-500 datamängd, och två distinkta modeller har tagits fram. Den ena är en multivariat regressionsmodell för att förutspå exakta värden, och den andra modellen är en multivariat klassifierare som förutspår riktningen på nästa dags indexrörelse. Experimenten visar för den konfiguration som presenteras i rapporten att LSTM RNN inte passar för att förutspå exakta värden för indexet, men ger tillfredsställande resultat när modellen ska förutsäga indexets framtida riktning.

This thesis demonstrates an approach to predict the count of social unrest events in South Africa. A comparison is made between traditional forecasting approaches and neural networks; the traditional forecast method selected being the Autoregressive Integrated Moving Average (ARIMA model). The type of neural network implemented was the Long Short-Term Memory (LSTM) neural network. The basic theoretical concepts of ARIMA and LSTM neural networks are explained and subsequently, the patterns of the social unrest time series were analysed using time series exploratory techniques. The social unrest time series contained a significant number of irregular fluctuations with a non-linear trend. The structure of the social unrest time series suggested that traditional linear approaches would fail to model the non-linear behaviour of the time series. This thesis confirms this finding. Twelve experiments were conducted, and in these experiments, features, scaling procedures and model configurations are varied (i.e. univariate and multivariate models). Multivariate LSTM achieved the lowest forecast errors and performance improved as more explanatory features were introduced. The ARIMA model's performance deteriorated with added complexity and the univariate ARIMA produced lower forecast errors compared to the multivariate ARIMA. In conclusion, it can be claimed that multivariate LSTM neural networks are useful for predicting social unrest events.

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.

With the PSD2 open banking revolution FinTechs obtained a key role in the financial industry. This role implies the inquiry and development of new techniques, products and solutions to compete with other players in this area. The aim of this thesis is to investigate the applicability of the state-of-the-art Deep Learning techniques for Credit Risk Modeling. In order to accomplish it, a PSD2-related synthetic and anonymized dataset has been used to simulate an application process with only one account per user. Firstly, a machine-readable representation of the bank accounts has been created, starting from the raw transactions’ data and scaling the variables using the quantile function. Afterwards, a Deep Neural Network has been created in order to capture the complex relations between the input variables and to extract information from the accounts’ representations. The proposed architecture accomplished the assigned tasks with a Gini index of 0.55, exploiting a Convolutional encoder to extract features from the inputs and a Recurrent decoder to analyze them.

Anomaly detection in satellite has not been well-documented due to the unavailability of satellite data, while it becomes more and more important with the increasing popularity of satellite applications. Our work focus on the anomaly detection by prediction on the dataset from the satellite, where we try and compare performance among recurrent neural network (RNN), Long Short-Term Memory (LSTM) and conventional neural network (NN). We conclude that LSTM with input length p=16, dimensionality n=32, output length q=2, 128 neurons and without maximum overlap is the best in terms of balanced accuracy. And LSTM with p=128, n=32, q=16, 128 and without maximum overlap outperforms most with respect to AUC metric. We also invent award function as a new performance metric trying to capture not only the correctness of decisions that NN made but also the amount of confidence in making its decisions, and we propose two candidates of award function. Regrettably, they partially meet our expectation as they possess a fatal defect which has been proved both from practical and theoretical viewpoints.

Neural network approaches have become popular in the field of automatic speech recognition (ASR). Most ASR methods use audio data to classify words. Lip reading ASR techniques utilize only video data, which compensates for noisy environments where audio may be compromised. A comprehensive approach, including the vetting of datasets and development of a preprocessing chain, to video-based ASR is developed. This approach will be based on neural networks, namely 3D convolutional neural networks (3D-CNN) and Long short-term memory (LSTM). These types of neural networks are designed to take in temporal data such as videos. Various combinations of different neural network architecture and preprocessing techniques are explored. The best performing neural network architecture, a CNN with bidirectional LSTM, compares favorably against recent works on video-based ASR.