Language is the principal medium for ideas, while dialogue is the most natural and effective way for humans to interact with and access information from machines. Natural language generation (NLG) is a critical component of spoken dialogue and it has a significant impact on usability and perceived quality. Many commonly used NLG systems employ rules and heuristics, which tend to generate inflexible and stylised responses without the natural variation of human language. However, the frequent repetition of identical output forms can quickly make dialogue become tedious for most real-world users. Additionally, these rules and heuristics are not scalable and hence not trivially extensible to other domains or languages. A statistical approach to language generation can learn language decisions directly from data without relying on hand-coded rules or heuristics, which brings scalability and flexibility to NLG. Statistical models also provide an opportunity to learn in-domain human colloquialisms and cross-domain model adaptations. A robust and quasi-supervised NLG model is proposed in this thesis. The model leverages a Recurrent Neural Network (RNN)-based surface realiser and a gating mechanism applied to input semantics. The model is motivated by the Long-Short Term Memory (LSTM) network. The RNN-based surface realiser and gating mechanism use a neural network to learn end-to-end language generation decisions from input dialogue act and sentence pairs; it also integrates sentence planning and surface realisation into a single optimisation problem. The single optimisation not only bypasses the costly intermediate linguistic annotations but also generates more natural and human-like responses. Furthermore, a domain adaptation study shows that the proposed model can be readily adapted and extended to new dialogue domains via a proposed recipe. Continuing the success of end-to-end learning, the second part of the thesis speculates on building an end-to-end dialogue system by framing it as a conditional generation problem. The proposed model encapsulates a belief tracker with a minimal state representation and a generator that takes the dialogue context to produce responses. These features suggest comprehension and fast learning. The proposed model is capable of understanding requests and accomplishing tasks after training on only a few hundred human-human dialogues. A complementary Wizard-of-Oz data collection method is also introduced to facilitate the collection of human-human conversations from online workers. The results demonstrate that the proposed model can talk to human judges naturally, without any difficulty, for a sample application domain. In addition, the results also suggest that the introduction of a stochastic latent variable can help the system model intrinsic variation in communicative intention much better.

The present work focuses on the two areas of investigation: system identification of helicopter and design of controller for the helicopter. Helicopter system identification, the first subject of investigation in this thesis, can be described as the extraction of system characteristics/dynamics from measured flight test data. Wind tunnel experimental data suffers from scale effects and model deficiencies. The increasing need for accurate models for the design of high bandwidth control system for helicopters has initiated a renewed interest in and a more active use of system identification. Besides, system identification is likely to become mandatory in the future for model validation of ground based helicopter simulators. Such simulators require accurate models in order to be accepted by pilots and regulatory authorities like Federal Aviation Regulation for realistic complementary helicopter mission training. Two approaches are widely used for system identification, namely, black box and gray box approach. In the black-box approach, the relationship between input-output data is approximated using nonparametric methods such as neural networks and in such a case, internal details of the system and model structure may not be known. In the gray box approach, parameters are estimated after defining the model structure. In this thesis, both black box and gray box approaches are investigated. In the black box approach, in this thesis, a comparative study and analysis of different Recurrent Neural Networks(RNN) for the identification of helicopter dynamics using flight data is investigated. Three different RNN architectures namely, Nonlinear Auto Regressive eXogenous input(NARX) model, neural network with internal memory known as Memory Neuron Networks(MNN)and Recurrent MultiLayer perceptron (RMLP) networks are used to identify dynamics of the helicopter at various flight conditions. Based on the results, the practical utility, advantages and limitations of the three models are critically appraised and it is found that the NARX model is most suitable for the identification of helicopter dynamics. In the gray box approach, helicopter model parameters are estimated after defining the model structure. The identification process becomes more difficult as the number of degrees-of-freedom and model parameters increase. To avoid the drawbacks of conventional methods, neural network based techniques, called the delta method is investigated in this thesis. This method does not require initial estimates of the parameters and the parameters can be directly extracted from the flight data. The Radial Basis Function Network(RBFN)is used for the purpose of estimation of parameters. It is shown that RBFN is able to satisfactorily estimate stability and control derivatives using the delta method. The second area of investigation addressed in this thesis is the control of helicopter in flight. Helicopter requires use of a control system to achieve satisfactory flight. Designing a classical controller involves developing a nonlinear model of the helicopter and extracting linearized state space matrices from the nonlinear model at various flight conditions. After examining the stability characteristics of the helicopter, the desired response is obtained using a feedback control system. The scheduling of controller gains over the entire envelope is used to obtain the desired response. In the present work, a helicopter having a soft inplane four bladed hingeless main rotor and a four-bladed tail rotor with conventional mechanical controls is considered. For this helicopter, a mathematical model and also a model based on neural network (using flight data) has been developed. As a precursor, a feed back controller, the Stability Augmentation System(SAS), is designed using linear quadratic regulator control with full state feedback and LQR with out put feedback approaches. SAS is designed to meet the handling qualities specification known as Aeronautical Design Standard ADS-33E-PRF. The control gains have been tuned with respect to forward speed and gain scheduling has been arrived at. The SAS in the longitudinal axis meets the requirement of the Level1 handling quality specifications in hover and low speed as well as for forward speed flight conditions. The SAS in the lateral axis meets the requirement of the Level2 handling quality specifications in both hover and low speed as well as for forward speed flight conditions. Such conventional design of control has served useful purposes, however, it requires considerable flight testing which is time consuming, to demonstrate and tune these control law gains. In modern helicopters, the stringent requirements and non-linear maneuvers make the controller design further complicated. Hence, new design tools have to be explored to control such helicopters. Among the many approaches in adaptive control, neural networks present a potential alternative for modeling and control of nonlinear dynamical systems due to their approximating capabilities and inherent adaptive features. Furthermore, from a practical perspective, the massive parallelism and fast adaptability of neural network implementations provide more incentive for further investigation in problems involving dynamical systems with unknown non-linearity. Therefore, adaptive control approach based on neural networks is proposed in this thesis. A neural network based Feedback Error Neural adaptive Controller(FENC) is designed for a helicopter. The proposed controller scheme is based on feedback error learning strategy in which the outer loop neural controller enhances the inner loop conventional controller by compensating for unknown non-linearity and parameter un-certainties. Nonlinear Auto Regressive eXogenous input(NARX)neural network architecture is used to approximate the control law and the controller network parameters are adapted using updated rules Lyapunov synthesis. An offline (finite time interval)and on-line adaptation strategy is used to approximate system uncertainties. The results are validated using simulation studies on helicopter undergoing an agile maneuver. The study shows that the neuro-controller meets the requirements of ADS-33 handling quality specifications. Even though the tracking error is less in FENC scheme, the control effort required to follow the command is very high. To overcome these problems, a Direct Adaptive Neural Control(DANC)scheme to track the rate command signal is presented. The neural controller is designed to track rate command signal generated using the reference model. For the simulation study, a linearized helicopter model at different straight and level flight conditions is considered. A neural network with a linear filter architecture trained using back propagation through time is used to approximate the control law. The controller network parameters are adapted using updated rules Lyapunov synthesis. The off-line trained (for finite time interval)network provides the necessary stability and tracking performance. The on-line learning is used to adapt the network under varying flight conditions. The on-line learning ability is demonstrated through parameter uncertainties. The performance of the proposed direct adaptive neural controller is compared with feedback error learning neural controller. The performance of the controller has been validated at various flight conditions. The theoretical results are validated using simulation studies based on a nonlinear six degree-of-freedom helicopter undergoing an agile maneuver. Realistic gust and sensor noise are added to the system to study the disturbance rejection properties of the neural controllers. To investigate the on-line learning ability of the proposed neural controller, different fault scenarios representing large model error and control surface loss are considered. The performances of the proposed DANC scheme is compared with the FENC scheme. The study shows that the neuro-controller meets the requirements of ADS-33 handling quality specifications.

The aim of this work is to create a tool for automatic transcription of historical documents. The work is mainly focused on the recognition of texts from the period of modern times written using font Fraktur. The problem is solved with a newly designed recurrent convolutional neural networks and a Spatial Transformer Network. Part of the solution is also an implemented generator of artificial historical texts. Using this generator, an artificial data set is created on which the convolutional neural network for line recognition is trained. This network is then tested on real historical lines of text on which the network achieves up to 89.0 % of character accuracy. The contribution of this work is primarily the newly designed neural network for text line recognition and the implemented artificial text generator, with which it is possible to train the neural network to recognize real historical lines of text.

Optical character recognition (OCR) has been a topic of interest for many years. It is defined as the process of digitizing a document image into a sequence of characters. Despite decades of intense research, OCR systems with capabilities to that of human still remains an open challenge. In this work there is presented a design and implementation of such system, which is capable of detecting texts in graphical user interfaces.

Optical character recognition (OCR) has been a topic of interest for many years. It is defined as the process of digitizing a document image into a sequence of characters. Despite decades of intense research, OCR systems with capabilities to that of human still remains an open challenge. In this work there is presented a design and implementation of such system, which is capable of detecting texts in graphical user interfaces.

This master thesis deals with the implementation of various types of recurrent neural networks via programming language lua using torch library. It focuses on finding optimal strategy for training recurrent neural networks and also tries to minimize the duration of the training. Furthermore various types of regularization techniques are investigated and implemented into the recurrent neural network architecture. Implemented recurrent neural networks are compared on the speech recognition task using AMI dataset, where they model the acustic information. Their performance is also compared to standard feedforward neural network. Best results are achieved using BLSTM architecture. The recurrent neural network are also trained via CTC objective function on the TIMIT dataset. Best result is again achieved using BLSTM architecture.

Synthetic short positions constructed by equity options and stock loan short sells are linked by arbitrage. This thesis analyses the link by considering the implied volatility surface (IVS) at 80%, 100%, and 120% moneyness, and stock loan variables such as benchmark rate (rt), utilization, short interest, and transaction trends to inspect time-dependent structures between the two assets. By applying multiple multivariate time-series analyses in terms of vector autoregression (VAR) and the recurrent neural networks long short-term memory (LSTM) and gated recurrent units (GRU) with a sliding window methodology. This thesis discovers linear and complex relationships between the IVS and stock loan data. The three-day-ahead out-of-sample LSTM forecast of IV at 80% moneyness improved by including lagged values of rt and yielded 19.6% MAPE and forecasted correct direction 81.1% of samples. The corresponding 100% moneyness GRU forecast was also improved by including stock loan data, at 10.8% MAPE and correct directions for 60.0% of samples. The 120% moneyness VAR forecast did not improve with stock loan data at 26.5% MAPE and correct directions for 66.2% samples. The one-month-ahead rt VAR forecast improved by including a lagged IVS, at 25.5% MAPE and 63.6% correct directions. The presented data was optimal for each target variable, showing that the application of LSTM and GRU was justified. These results indicate that considering stock loan data when forecasting IVS for 80% and 100% moneyness is advised to gain exploitable insights for short-term positions. They are further validated since the different models yielded parallel inferences. Similar analysis with other equity is advised to gain insights into the relationship and improve such forecasts. / Syntetiska kortpositioner konstruerade av aktieoptioner och blankning med aktielån är kopplade med arbitrage. Denna tes analyserar kopplingen genom att överväga den implicerade volatilitetsytan vid 80%, 100% och 120% moneyness och aktielånvariabler såsom referensränta rt, låneutnyttjande, låneintresse, och transaktionstrender för att granska tidsberoende strukturer mellan de två tillgångarna. Genom att tillämpa multipel multidimensionell tidsserieanalys såsom vektorautoregression (VAR) och de rekursiva neurala nätverken long short-term memory (LSTM) och gated recurrent units (GRU). Tesen upptäcker linjära och komplexa samband mellan implicerade volatilitetsytor och aktielånedata. Tre dagars LSTM-prognos av implicerade volatiliteten vid 80% moneyness förbättrades genom att inkludera fördröjda värden av rt och gav 19,6% MAPE och prognostiserade korrekt riktning för 81,1% av prover. Motsvarande 100% moneyness GRU-prognos förbättrades också genom att inkludera aktielånedata, resulterande i 10,8% MAPE och korrekt riktning för 60,0% av prover. VAR-prognosen för 120% moneyness förbättrades inte med alternativa data på 26,5% MAPE och korrekt riktning för 66,2% av prover. En månads VAR-prognos för rt förbättrades genom att inkludera en fördröjd implicerad volatilitetsyta, resulterande i 25,5% MAPE och 63,6% korrekta riktningar. Presenterad statistik var optimala för dessa variabler, vilket visar att tillämpningen av LSTM och GRU var motiverad. Därav rekommenderas det att inkludera aktielånedata för prognostisering av implicerade volatilitetsytor för 80% och 100% moneyness, speciellt för kortsiktiga positioner. Resultaten valideras ytterligare eftersom de olika modellerna gav dylika slutsatser. Liknande analys med andra aktier är rekommenderat för att få insikter i förhållandet och förbättra sådana prognoser.

The development of machine learning research has provided statistical innovations and further developments within the field of time series analysis. This study seeks to investigate two different approaches on artificial neural network models based on different learning techniques, and answering how well the neural network approach compares with a basic autoregressive approach, as well as how the artificial neural network models compare to each other. The models were compared and analyzed in regards to the univariate forecast accuracy on 20 randomly drawn time series from two different time frequencies from the M4 competition dataset. Forecasting was made dependent on one time lag (t-1) and forecasted three and six steps ahead respectively. The artificial neural network models outperformed the baseline Autoregressive model, showing notably lower mean average percentage error overall. The Multilayered perceptron models performed better than the Long short-term memory model overall, whereas the Long short-term memory model showed improvement on longer prediction time dimensions. As the training were done univariately  on a limited set of time steps, it is believed that the one layered-approach gave a good enough approximation on the data, whereas the added layer couldn’t fully utilize its strengths of processing power. Likewise, the Long short-term memory model couldn’t fully demonstrate the advantagements of recurrent learning. Using the same dataset, further studies could be made with another approach to data processing. Implementing an unsupervised approach of clustering the data before analysis, the same models could be tested with multivariate analysis on models trained on multiple time series simultaneously.

As the complexity of deep learning model increases, the transparency of the systems does the opposite. It may be hard to understand the predictions a deep learning model makes, but even harder to understand why these predictions are made. Using eXplainable AI (XAI), we can gain greater knowledge of how the model operates and how the input in which the model receives can change its predictions. In this thesis, we apply Integrated Gradients (IG), an XAI method primarily used on image data and on datasets containing tabular and time-series data. We also evaluate how the results of IG differ from various types of models and how the change of baseline can change the outcome. In these results, we observe that IG can be applied to both sequenced and nonsequenced data, with varying results. We can see that the gradient baseline does not affect the results of IG on models such as RNN, LSTM, and GRU, where the data contains time series, as much as it does for models like MLP with nonsequenced data. To confirm this, we also applied IG to SVM models, which gave the results that the choice of gradient baseline has a significant impact on the results of IG.

<p>    </p> <p>In order to advance next-generation communication systems, it is critical to enhance the state-of-the-art communication architectures, such as device-to-device (D2D), multiple- input multiple-output (MIMO), and intelligent reflecting surface (IRS), in terms of achieving high data rate, low latency, and high energy efficiency. In the first part of this dissertation, we address joint learning and optimization methodologies on cutting-edge network archi- tectures. First, we consider D2D networks equipped with MIMO systems. In particular, we address the problem of minimizing the network overhead in D2D networks, defined as the sum of time and energy required for processing tasks at devices, through the design for MIMO beamforming and communication/computation resource allocation. Second, we address IRS-assisted communication systems. Specifically, we study an adaptive IRS control scheme considering realistic IRS reflection behavior and channel environments, and propose a novel adaptive codebook-based limited feedback protocol and learning-based solutions for codebook updates. </p> <p><br></p> <p>Furthermore, in order for revolutionary innovations to emerge for future generations of communications, it is crucial to explore and address fundamental, long-standing open problems for communications, such as the design of practical codes for a variety of important channel models. In the later part of this dissertation, we study the design of practical codes for feedback-enabled communication channels, i.e., feedback codes. The existing feedback codes, which have been developed over the past six decades, have been demonstrated to be vulnerable to high forward/feedback noises, due to the non-triviality of the design of feedback codes. We propose a novel recurrent neural network (RNN) autoencoder-based architecture to mitigate the susceptibility to high channel noises by incorporating domain knowledge into the design of the deep learning architecture. Using this architecture, we suggest a new class of non-linear feedback codes that increase robustness to forward/feedback noise in additive White Gaussian noise (AWGN) channels with feedback. </p>