Pretraining Convolutional Neural Networks for Visual Recognition

Convolutional Neural Networks (CNN) have become the state-of-the-art methods on many large scale visual recognition tasks. For a lot of practical applications, CNN architectures have a restrictive requirement: A huge amount of labeled data are needed for training. The idea of generative pretraining is to obtain initial weights of the network by training the network in a completely unsupervised way and then fine-tune the weights for the task at hand using supervised learning. In this thesis, a general introduction to Deep Neural Networks and algorithms are given and these methods are applied to classification tasks of handwritten digits and natural images for developing unsupervised feature learning. The goal of this thesis is to find out if the effect of pretraining is damped by recent practical advances in optimization and regularization of CNN. The experimental results show that pretraining is still a substantial regularizer, however, not a necessary step in training Convolutional Neural Networks with rectified activations. On handwritten digits, the proposed pretraining model achieved a classification accuracy comparable to the state-of-the-art methods.

Modelling non-industrial private forest landowners' strategic decision making by using logistic regression and neural networks: case of predicting the choice of forest taxation basis

Summary

On the differential evolution algorithm and its appliation to training radial basis function networks

The parameter setting of a differential evolution algorithm must meet several requirements: efficiency, effectiveness, and reliability. Problems vary. The solution of a particular problem can be represented in different ways. An algorithm most efficient in dealing with a particular representation may be less efficient in dealing with other representations. The development of differential evolution-based methods contributes substantially to research on evolutionary computing and global optimization in general. The objective of this study is to investigatethe differential evolution algorithm, the intelligent adjustment of its controlparameters, and its application. In the thesis, the differential evolution algorithm is first examined using different parameter settings and test functions. Fuzzy control is then employed to make control parameters adaptive based on an optimization process and expert knowledge. The developed algorithms are applied to training radial basis function networks for function approximation with possible variables including centers, widths, and weights of basis functions and both having control parameters kept fixed and adjusted by fuzzy controller. After the influence of control variables on the performance of the differential evolution algorithm was explored, an adaptive version of the differential evolution algorithm was developed and the differential evolution-based radial basis function network training approaches were proposed. Experimental results showed that the performance of the differential evolution algorithm is sensitive to parameter setting, and the best setting was found to be problem dependent. The fuzzy adaptive differential evolution algorithm releases the user load of parameter setting and performs better than those using all fixedparameters. Differential evolution-based approaches are effective for training Gaussian radial basis function networks.

Artificially Intelligent and Adaptive Methods for Prediction and Analysis of Superheater Fireside Corrosion in Fluidized Bed Boilers

Superheater corrosion causes vast annual losses for the power companies. With a reliable corrosion prediction method, the plants can be designed accordingly, and knowledge of fuel selection and determination of process conditions may be utilized to minimize superheater corrosion. Growing interest to use recycled fuels creates additional demands for the prediction of corrosion potential. Models depending on corrosion theories will fail, if relations between the inputs and the output are poorly known. A prediction model based on fuzzy logic and an artificial neural network is able to improve its performance as the amount of data increases. The corrosion rate of a superheater material can most reliably be detected with a test done in a test combustor or in a commercial boiler. The steel samples can be located in a special, temperature-controlled probe, and exposed to the corrosive environment for a desired time. These tests give information about the average corrosion potential in that environment. Samples may also be cut from superheaters during shutdowns. The analysis ofsamples taken from probes or superheaters after exposure to corrosive environment is a demanding task: if the corrosive contaminants can be reliably analyzed, the corrosion chemistry can be determined, and an estimate of the material lifetime can be given. In cases where the reason for corrosion is not clear, the determination of the corrosion chemistry and the lifetime estimation is more demanding. In order to provide a laboratory tool for the analysis and prediction, a newapproach was chosen. During this study, the following tools were generated: · Amodel for the prediction of superheater fireside corrosion, based on fuzzy logic and an artificial neural network, build upon a corrosion database developed offuel and bed material analyses, and measured corrosion data. The developed model predicts superheater corrosion with high accuracy at the early stages of a project. · An adaptive corrosion analysis tool based on image analysis, constructedas an expert system. This system utilizes implementation of user-defined algorithms, which allows the development of an artificially intelligent system for thetask. According to the results of the analyses, several new rules were developed for the determination of the degree and type of corrosion. By combining these two tools, a user-friendly expert system for the prediction and analyses of superheater fireside corrosion was developed. This tool may also be used for the minimization of corrosion risks by the design of fluidized bed boilers.

On Modelling, System Identification and Control of Servo-Systems with a Flexible Load

The present study was done with two different servo-systems. In the first system, a servo-hydraulic system was identified and then controlled by a fuzzy gainscheduling controller. The second servo-system, an electro-magnetic linear motor in suppressing the mechanical vibration and position tracking of a reference model are studied by using a neural network and an adaptive backstepping controller respectively. Followings are some descriptions of research methods. Electro Hydraulic Servo Systems (EHSS) are commonly used in industry. These kinds of systems are nonlinearin nature and their dynamic equations have several unknown parameters.System identification is a prerequisite to analysis of a dynamic system. One of the most promising novel evolutionary algorithms is the Differential Evolution (DE) for solving global optimization problems. In the study, the DE algorithm is proposed for handling nonlinear constraint functionswith boundary limits of variables to find the best parameters of a servo-hydraulic system with flexible load. The DE guarantees fast speed convergence and accurate solutions regardless the initial conditions of parameters. The control of hydraulic servo-systems has been the focus ofintense research over the past decades. These kinds of systems are nonlinear in nature and generally difficult to control. Since changing system parameters using the same gains will cause overshoot or even loss of system stability. The highly non-linear behaviour of these devices makes them ideal subjects for applying different types of sophisticated controllers. The study is concerned with a second order model reference to positioning control of a flexible load servo-hydraulic system using fuzzy gainscheduling. In the present research, to compensate the lack of dampingin a hydraulic system, an acceleration feedback was used. To compare the results, a pcontroller with feed-forward acceleration and different gains in extension and retraction is used. The design procedure for the controller and experimental results are discussed. The results suggest that using the fuzzy gain-scheduling controller decrease the error of position reference tracking. The second part of research was done on a PermanentMagnet Linear Synchronous Motor (PMLSM). In this study, a recurrent neural network compensator for suppressing mechanical vibration in PMLSM with a flexible load is studied. The linear motor is controlled by a conventional PI velocity controller, and the vibration of the flexible mechanism is suppressed by using a hybrid recurrent neural network. The differential evolution strategy and Kalman filter method are used to avoid the local minimum problem, and estimate the states of system respectively. The proposed control method is firstly designed by using non-linear simulation model built in Matlab Simulink and then implemented in practical test rig. The proposed method works satisfactorily and suppresses the vibration successfully. In the last part of research, a nonlinear load control method is developed and implemented for a PMLSM with a flexible load. The purpose of the controller is to track a flexible load to the desired position reference as fast as possible and without awkward oscillation. The control method is based on an adaptive backstepping algorithm whose stability is ensured by the Lyapunov stability theorem. The states of the system needed in the controller are estimated by using the Kalman filter. The proposed controller is implemented and tested in a linear motor test drive and responses are presented.

Application of Neural Network Methods in the Prediction of Stock Markets

The purpose of the research is to define practical profit which can be achieved using neural network methods as a prediction instrument. The thesis investigates the ability of neural networks to forecast future events. This capability is checked on the example of price prediction during intraday trading on stock market. The executed experiments show predictions of average 1, 2, 5 and 10 minutes’ prices based on data of one day and made by two different types of forecasting systems. These systems are based on the recurrent neural networks and back propagation neural nets. The precision of the predictions is controlled by the absolute error and the error of market direction. The economical effectiveness is estimated by a special trading system. In conclusion, the best structures of neural nets are tested with data of 31 days’ interval. The best results of the average percent of profit from one transaction (buying + selling) are 0.06668654, 0.188299453, 0.349854787 and 0.453178626, they were achieved for prediction periods 1, 2, 5 and 10 minutes. The investigation can be interesting for the investors who have access to a fast information channel with a possibility of every-minute data refreshment.

On the Analysis and Control of a Linear Synchronous Servomotor with a Flexible Load

Industry's growing need for higher productivity is placing new demands on mechanisms connected with electrical motors, because these can easily lead to vibration problems due to fast dynamics. Furthermore, the nonlinear effects caused by a motor frequently reduce servo stability, which diminishes the controller's ability to predict and maintain speed. Hence, the flexibility of a mechanism and its control has become an important area of research. The basic approach in control system engineering is to assume that the mechanism connected to a motor is rigid, so that vibrations in the tool mechanism, reel, gripper or any apparatus connected to the motor are not taken into account. This might reduce the ability of the machine system to carry out its assignment and shorten the lifetime of the equipment. Nonetheless, it is usually more important to know how the mechanism, or in other words the load on the motor, behaves. A nonlinear load control method for a permanent magnet linear synchronous motor is developed and implemented in the thesis. The purpose of the controller is to track a flexible load to the desired velocity reference as fast as possible and without awkward oscillations. The control method is based on an adaptive backstepping algorithm with its stability ensured by the Lyapunov stability theorem. As a reference controller for the backstepping method, a hybrid neural controller is introduced in which the linear motor itself is controlled by a conventional PI velocity controller and the vibration of the associated flexible mechanism is suppressed from an outer control loop using a compensation signal from a multilayer perceptron network. To avoid the local minimum problem entailed in neural networks, the initial weights are searched for offline by means of a differential evolution algorithm. The states of a mechanical system for controllers are estimated using the Kalman filter. The theoretical results obtained from the control design are validated with the lumped mass model for a mechanism. Generalization of the mechanism allows the methods derived here to be widely implemented in machine automation. The control algorithms are first designed in a specially introduced nonlinear simulation model and then implemented in the physical linear motor using a DSP (Digital Signal Processor) application. The measurements prove that both controllers are capable of suppressing vibration, but that the backstepping method is superior to others due to its accuracy of response and stability properties.

The Connectivity of Musical Aptitude and Foreign Language Learning Skills: Neural and Behavioural Evidence

Given the structural and acoustical similarities between speech and music, and possible overlapping cerebral structures in speech and music processing, a possible relationship between musical aptitude and linguistic abilities, especially in terms of second language pronunciation skills, was investigated. Moreover, the laterality effect of the mother tongue was examined with both adults and children by means of dichotic listening scores. Finally, two event-related potential studies sought to reveal whether children with advanced second language pronunciation skills and higher general musical aptitude differed from children with less-advanced pronunciation skills and less musical aptitude in accuracy when preattentively processing mistuned triads and music / speech sound durations. The results showed a significant relationship between musical aptitude, English language pronunciation skills, chord discrimination ability, and sound-change-evoked brain activation in response to musical stimuli (durational differences and triad contrasts). Regular music practice may also have a modulatory effect on the brain’s linguistic organization and cause altered hemispheric functioning in those who have regularly practised music for years. Based on the present results, it is proposed that language skills, both in production and discrimination, are interconnected with perceptual musical skills.

Use of neural networks in robot positioning of large flexible redundant manipulators

Deflection compensation of flexible boom structures in robot positioning is usually done using tables containing the magnitude of the deflection with inverse kinematics solutions of a rigid structure. The number of table values increases greatly if the working area of the boom is large and the required positioning accuracy is high. The inverse kinematics problems are very nonlinear, and if the structure is redundant, in some cases it cannot be solved in a closed form. If the structural flexibility of the manipulator arms is taken into account, the problem is almost impossible to solve using analytical methods. Neural networks offer a possibility to approximate any linear or nonlinear function. This study presents four different methods of using neural networks in the static deflection compensation and inverse kinematics solution of a flexible hydraulically driven manipulator. The training information required for training neural networks is obtained by employing a simulation model that includes elasticity characteristics. The functionality of the presented methods is tested based on the simulated and measured results of positioning accuracy. The simulated positioning accuracy is tested in 25 separate coordinate points. For each point, the positioning is tested with five different mass loads. The mean positioning error of a manipulator decreased from 31.9 mm to 4.1 mm in the test points. This accuracy enables the use of flexible manipulators in the positioning of larger objects. The measured positioning accuracy is tested in 9 separate points using three different mass loads. The mean positioning error decreased from 10.6 mm to 4.7 mm and the maximum error from 27.5 mm to 11.0 mm.

MCMC Analysis Of Classical Time Series Algorithms

Identification of order of an Autoregressive Moving Average Model (ARMA) by the usual graphical method is subjective. Hence, there is a need of developing a technique to identify the order without employing the graphical investigation of series autocorrelations. To avoid subjectivity, this thesis focuses on determining the order of the Autoregressive Moving Average Model using Reversible Jump Markov Chain Monte Carlo (RJMCMC). The RJMCMC selects the model from a set of the models suggested by better fitting, standard deviation errors and the frequency of accepted data. Together with deep analysis of the classical Box-Jenkins modeling methodology the integration with MCMC algorithms has been focused through parameter estimation and model fitting of ARMA models. This helps to verify how well the MCMC algorithms can treat the ARMA models, by comparing the results with graphical method. It has been seen that the MCMC produced better results than the classical time series approach.

Verbal learning in mild cognitive impairment and alzheimer's disease : behavioural and neural approaches

The main focus of the present thesis was at verbal episodic memory processes that are particularly vulnerable to preclinical and clinical Alzheimer’s disease (AD). Here these processes were studied by a word learning paradigm, cutting across the domains of memory and language learning studies. Moreover, the differentiation between normal aging, mild cognitive impairment (MCI) and AD was studied by the cognitive screening test CERAD. In study I, the aim was to examine how patients with amnestic MCI differ from healthy controls in the different CERAD subtests. Also, the sensitivity and specificity of the CERAD screening test to MCI and AD was examined, as previous studies on the sensitivity and specificity of the CERAD have not included MCI patients. The results indicated that MCI is characterized by an encoding deficit, as shown by the overall worse performance on the CERAD Wordlist learning test compared with controls. As a screening test, CERAD was not very sensitive to MCI. In study II, verbal learning and forgetting in amnestic MCI, AD and healthy elderly controls was investigated with an experimental word learning paradigm, where names of 40 unfamiliar objects (mainly archaic tools) were trained with or without semantic support. The object names were trained during a 4-day long period and a follow-up was conducted one week, 4 weeks and 8 weeks after the training period. Manipulation of semantic support was included in the paradigm because it was hypothesized that semantic support might have some beneficial effects in the present learning task especially for the MCI group, as semantic memory is quite well preserved in MCI in contrast to episodic memory. We found that word learning was significantly impaired in MCI and AD patients, whereas forgetting patterns were similar across groups. Semantic support showed a beneficial effect on object name retrieval in the MCI group 8 weeks after training, indicating that the MCI patients’ preserved semantic memory abilities compensated for their impaired episodic memory. The MCI group performed equally well as the controls in the tasks tapping incidental learning and recognition memory, whereas the AD group showed impairment. Both the MCI and the AD group benefited less from phonological cueing than the controls. Our findings indicate that acquisition is compromised in both MCI and AD, whereas long13 term retention is not affected to the same extent. Incidental learning and recognition memory seem to be well preserved in MCI. In studies III and IV, the neural correlates of naming newly learned objects were examined in healthy elderly subjects and in amnestic MCI patients by means of positron emission tomography (PET) right after the training period. The naming of newly learned objects by healthy elderly subjects recruited a left-lateralized network, including frontotemporal regions and the cerebellum, which was more extensive than the one related to the naming of familiar objects (study III). Semantic support showed no effects on the PET results for the healthy subjects. The observed activation increases may reflect lexicalsemantic and lexical-phonological retrieval, as well as more general associative memory mechanisms. In study IV, compared to the controls, the MCI patients showed increased anterior cingulate activation when naming newly learned objects that had been learned without semantic support. This suggests a recruitment of additional executive and attentional resources in the MCI group.

Neural correlates of language learning in adults

The human language-learning ability persists throughout life, indicating considerable flexibility at the cognitive and neural level. This ability spans from expanding the vocabulary in the mother tongue to acquisition of a new language with its lexicon and grammar. The present thesis consists of five studies that tap both of these aspects of adult language learning by using magnetoencephalography (MEG) and functional magnetic resonance imaging (fMRI) during language processing and language learning tasks. The thesis shows that learning novel phonological word forms, either in the native tongue or when exposed to a foreign phonology, activates the brain in similar ways. The results also show that novel native words readily become integrated in the mental lexicon. Several studies in the thesis highlight the left temporal cortex as an important brain region in learning and accessing phonological forms. Incidental learning of foreign phonological word forms was reflected in functionally distinct temporal lobe areas that, respectively, reflected short-term memory processes and more stable learning that persisted to the next day. In a study where explicitly trained items were tracked for ten months, it was found that enhanced naming-related temporal and frontal activation one week after learning was predictive of good long-term memory. The results suggest that memory maintenance is an active process that depends on mechanisms of reconsolidation, and that these process vary considerably between individuals. The thesis put special emphasis on studying language learning in the context of language production. The neural foundation of language production has been studied considerably less than that of perceptive language, especially on the sentence level. A well-known paradigm in language production studies is picture naming, also used as a clinical tool in neuropsychology. This thesis shows that accessing the meaning and phonological form of a depicted object are subserved by different neural implementations. Moreover, a comparison between action and object naming from identical images indicated that the grammatical class of the retrieved word (verb, noun) is less important than the visual content of the image. In the present thesis, the picture naming was further modified into a novel paradigm in order to probe sentence-level speech production in a newly learned miniature language. Neural activity related to grammatical processing did not differ between the novel language and the mother tongue, but stronger neural activation for the novel language was observed during the planning of the upcoming output, likely related to more demanding lexical retrieval and short-term memory. In sum, the thesis aimed at examining language learning by combining different linguistic domains, such as phonology, semantics, and grammar, in a dynamic description of language processing in the human brain.

Neural mechanisms underlying conscious and unconscious vision. Evidence from event-related potentials and transcranial magnetic stimulation

Vision affords us with the ability to consciously see, and use this information in our behavior. While research has produced a detailed account of the function of the visual system, the neural processes that underlie conscious vision are still debated. One of the aims of the present thesis was to examine the time-course of the neuroelectrical processes that correlate with conscious vision. The second aim was to study the neural basis of unconscious vision, that is, situations where a stimulus that is not consciously perceived nevertheless influences behavior. According to current prevalent models of conscious vision, the activation of visual cortical areas is not, as such, sufficient for consciousness to emerge, although it might be sufficient for unconscious vision. Conscious vision is assumed to require reciprocal communication between cortical areas, but views differ substantially on the extent of this recurrent communication. Visual consciousness has been proposed to emerge from recurrent neural interactions within the visual system, while other models claim that more widespread cortical activation is needed for consciousness. Studies I-III compared models of conscious vision by studying event-related potentials (ERP). ERPs represent the brain’s average electrical response to stimulation. The results support the model that associates conscious vision with activity localized in the ventral visual cortex. The timing of this activity corresponds to an intermediate stage in visual processing. Earlier stages of visual processing may influence what becomes conscious, although these processes do not directly enable visual consciousness. Late processing stages, when more widespread cortical areas are activated, reflect the access to and manipulation of contents of consciousness. Studies IV and V concentrated on unconscious vision. By using transcranial magnetic stimulation (TMS) we show that when early visual cortical processing is disturbed so that subjects fail to consciously perceive visual stimuli, they may nevertheless guess (above chance-level) the location where the visual stimuli were presented. However, the results also suggest that in a similar situation, early visual cortex is necessary for both conscious and unconscious perception of chromatic information (i.e. color). Chromatic information that remains unconscious may influence behavioral responses when activity in visual cortex is not disturbed by TMS. Our results support the view that early stimulus-driven (feedforward) activation may be sufficient for unconscious processing. In conclusion, the results of this thesis support the view that conscious vision is enabled by a series of processing stages. The processes that most closely correlate with conscious vision take place in the ventral visual cortex ~200 ms after stimulus presentation, although preceding time-periods and contributions from other cortical areas such as the parietal cortex are also indispensable. Unconscious vision relies on intact early visual activation, although the location of visual stimulus may be unconsciously resolved even when activity in the early visual cortex is interfered with.