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.

Deep convolutional neural networks for semantic video object segmentation

In this thesis, we propose to infer pixel-level labelling in video by utilising only object category information, exploiting the intrinsic structure of video data. Our motivation is the observation that image-level labels are much more easily to be acquired than pixel-level labels, and it is natural to find a link between the image level recognition and pixel level classification in video data, which would transfer learned recognition models from one domain to the other one. To this end, this thesis proposes two domain adaptation approaches to adapt the deep convolutional neural network (CNN) image recognition model trained from labelled image data to the target domain exploiting both semantic evidence learned from CNN, and the intrinsic structures of unlabelled video data. Our proposed approaches explicitly model and compensate for the domain adaptation from the source domain to the target domain which in turn underpins a robust semantic object segmentation method for natural videos. We demonstrate the superior performance of our methods by presenting extensive evaluations on challenging datasets comparing with the state-of-the-art methods.

Evaluation of Machine Learning Classifiers for Mobile Network Intrusion Detection Systems

Mobile malwares are increasing with the growing number of Mobile users. Mobile malwares can perform several operations which lead to cybersecurity threats such as, stealing financial or personal information, installing malicious applications, sending premium SMS, creating backdoors, keylogging and crypto-ransomware attacks. Knowing the fact that there are many illegitimate Applications available on the App stores, most of the mobile users remain careless about the security of their Mobile devices and become the potential victim of these threats. Previous studies have shown that not every antivirus is capable of detecting all the threats; due to the fact that Mobile malwares use advance techniques to avoid detection. A Network-based IDS at the operator side will bring an extra layer of security to the subscribers and can detect many advanced threats by analyzing their traffic patterns. Machine Learning(ML) will provide the ability to these systems to detect unknown threats for which signatures are not yet known. This research is focused on the evaluation of Machine Learning classifiers in Network-based Intrusion detection systems for Mobile Networks. In this study, different techniques of Network-based intrusion detection with their advantages, disadvantages and state of the art in Hybrid solutions are discussed. Finally, a ML based NIDS is proposed which will work as a subsystem, to Network-based IDS deployed by Mobile Operators, that can help in detecting unknown threats and reducing false positives. In this research, several ML classifiers were implemented and evaluated. This study is focused on Android-based malwares, as Android is the most popular OS among users, hence most targeted by cyber criminals. Supervised ML algorithms based classifiers were built using the dataset which contained the labeled instances of relevant features. These features were extracted from the traffic generated by samples of several malware families and benign applications. These classifiers were able to detect malicious traffic patterns with the TPR upto 99.6% during Cross-validation test. Also, several experiments were conducted to detect unknown malware traffic and to detect false positives. These classifiers were able to detect unknown threats with the Accuracy of 97.5%. These classifiers could be integrated with current NIDS', which use signatures, statistical or knowledge-based techniques to detect malicious traffic. Technique to integrate the output from ML classifier with traditional NIDS is discussed and proposed for future work.

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.

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.

Study of ligand-based virtual screening tools in computer-aided drug design

Virtual screening is a central technique in drug discovery today. Millions of molecules can be tested in silico with the aim to only select the most promising and test them experimentally. The topic of this thesis is ligand-based virtual screening tools which take existing active molecules as starting point for finding new drug candidates. One goal of this thesis was to build a model that gives the probability that two molecules are biologically similar as function of one or more chemical similarity scores. Another important goal was to evaluate how well different ligand-based virtual screening tools are able to distinguish active molecules from inactives. One more criterion set for the virtual screening tools was their applicability in scaffold-hopping, i.e. finding new active chemotypes. In the first part of the work, a link was defined between the abstract chemical similarity score given by a screening tool and the probability that the two molecules are biologically similar. These results help to decide objectively which virtual screening hits to test experimentally. The work also resulted in a new type of data fusion method when using two or more tools. In the second part, five ligand-based virtual screening tools were evaluated and their performance was found to be generally poor. Three reasons for this were proposed: false negatives in the benchmark sets, active molecules that do not share the binding mode, and activity cliffs. In the third part of the study, a novel visualization and quantification method is presented for evaluation of the scaffold-hopping ability of virtual screening tools.

Modelling the colour difference by artificial neural network

This thesis work studies the modelling of the colour difference using artificial neural network. Multilayer percepton (MLP) network is proposed to model CIEDE2000 colour difference formula. MLP is applied to classify colour points in CIE xy chromaticity diagram. In this context, the evaluation was performed using Munsell colour data and MacAdam colour discrimination ellipses. Moreover, in CIE xy chromaticity diagram just noticeable differences (JND) of MacAdam ellipses centres are computed by CIEDE2000, to compare JND of CIEDE2000 and MacAdam ellipses. CIEDE2000 changes the orientation of blue areas in CIE xy chromaticity diagram toward neutral areas, but on the whole it does not totally agree with the MacAdam ellipses. The proposed MLP for both modelling CIEDE2000 and classifying colour points showed good accuracy and achieved acceptable results.

Controlling full penetration in MAG welding by the application of infrared thermography and neural network

In this study, an infrared thermography based sensor was studied with regard to usability and the accuracy of sensor data as a weld penetration signal in gas metal arc welding. The object of the study was to evaluate a specific sensor type which measures thermography from solidified weld surface. The purpose of the study was to provide expert data for developing a sensor system in adaptive metal active gas (MAG) welding. Welding experiments with considered process variables and recorded thermal profiles were saved to a database for further analysis. To perform the analysis within a reasonable amount of experiments, the process parameter variables were gradually altered by at least 10 %. Later, the effects of process variables on weld penetration and thermography itself were considered. SFS-EN ISO 5817 standard (2014) was applied for classifying the quality of the experiments. As a final step, a neural network was taught based on the experiments. The experiments show that the studied thermography sensor and the neural network can be used for controlling full penetration though they have minor limitations, which are presented in results and discussion. The results are consistent with previous studies and experiments found in the literature.

Deep generative models for facial keypoints detection

A new area of machine learning research called deep learning, has moved machine learning closer to one of its original goals: artificial intelligence and general learning algorithm. The key idea is to pretrain models in completely unsupervised way and finally they can be fine-tuned for the task at hand using supervised learning. In this thesis, a general introduction to deep learning models and algorithms are given and these methods are applied to facial keypoints detection. The task is to predict the positions of 15 keypoints on grayscale face images. Each predicted keypoint is specified by an (x,y) real-valued pair in the space of pixel indices. In experiments, we pretrained deep belief networks (DBN) and finally performed a discriminative fine-tuning. We varied the depth and size of an architecture. We tested both deterministic and sampled hidden activations and the effect of additional unlabeled data on pretraining. The experimental results show that our model provides better results than publicly available benchmarks for the dataset.

Novel word learning ability in chronic post-stroke aphasia : variability and modality effects

Novel word learning has been rarely studied in people with aphasia (PWA), although it can provide a relatively pure measure of their learning potential, and thereby contribute to the development of effective aphasia treatment methods. The main aim of the present thesis was to explore the capacity of PWA for associative learning of word–referent pairings and cognitive-linguistic factors related to it. More specifically, the thesis examined learning and long-term maintenance of the learned pairings, the role of lexical-semantic abilities in learning as well as acquisition of phonological versus semantic information in associative novel word learning. Furthermore, the effect of modality on associative novel word learning and the neural underpinnings of successful learning were explored. The learning experiments utilized the Ancient Farming Equipment (AFE) paradigm that employs drawings of unfamiliar referents and their unfamiliar names. Case studies of Finnishand English-speaking people with chronic aphasia (n = 6) were conducted in the investigation. The learning results of PWA were compared to those of healthy control participants, and active production of the novel words and their semantic definitions was used as learning outcome measures. PWA learned novel word–novel referent pairings, but the variation between individuals was very wide, from more modest outcomes (Studies I–II) up to levels on a par with healthy individuals (Studies III–IV). In incidental learning of semantic definitions, none of the PWA reached the performance level of the healthy control participants. Some PWA maintained part of the learning outcomes up to months post-training, and one individual showed full maintenance of the novel words at six months post-training (Study IV). Intact lexical-semantic processing skills promoted learning in PWA (Studies I–II) but poor phonological short-term memory capacities did not rule out novel word learning. In two PWA with successful learning and long-term maintenance of novel word–novel referent pairings, learning relied on orthographic input while auditory input led to significantly inferior learning outcomes (Studies III–IV). In one of these individuals, this previously undetected modalityspecific learning ability was successfully translated into training with familiar but inaccessible everyday words (Study IV). Functional magnetic resonance imaging revealed that this individual had a disconnected dorsal speech processing pathway in the left hemisphere, but a right-hemispheric neural network mediated successful novel word learning via reading. Finally, the results of Study III suggested that the cognitive-linguistic profile may not always predict the optimal learning channel for an individual with aphasia. Small-scale learning probes seem therefore useful in revealing functional learning channels in post-stroke aphasia.

Socially shared metacognitive regulation during collaborative learning processes in student dyads and small groups

Traditionally metacognition has been theorised, methodologically studied and empirically tested from the standpoint mainly of individuals and their learning contexts. In this dissertation the emergence of metacognition is analysed more broadly. The aim of the dissertation was to explore socially shared metacognitive regulation (SSMR) as part of collaborative learning processes taking place in student dyads and small learning groups. The specific aims were to extend the concept of individual metacognition to SSMR, to develop methods to capture and analyse SSMR and to validate the usefulness of the concept of SSMR in two different learning contexts; in face-to-face student dyads solving mathematical word problems and also in small groups taking part in inquiry-based science learning in an asynchronous computer-supported collaborative learning (CSCL) environment. This dissertation is comprised of four studies. In Study I, the main aim was to explore if and how metacognition emerges during problem solving in student dyads and then to develop a method for analysing the social level of awareness, monitoring, and regulatory processes emerging during the problem solving. Two dyads comprised of 10-year-old students who were high-achieving especially in mathematical word problem solving and reading comprehension were involved in the study. An in-depth case analysis was conducted. Data consisted of over 16 (30–45 minutes) videotaped and transcribed face-to-face sessions. The dyads solved altogether 151 mathematical word problems of different difficulty levels in a game-format learning environment. The interaction flowchart was used in the analysis to uncover socially shared metacognition. Interviews (also stimulated recall interviews) were conducted in order to obtain further information about socially shared metacognition. The findings showed the emergence of metacognition in a collaborative learning context in a way that cannot solely be explained by individual conception. The concept of socially-shared metacognition (SSMR) was proposed. The results highlighted the emergence of socially shared metacognition specifically in problems where dyads encountered challenges. Small verbal and nonverbal signals between students also triggered the emergence of socially shared metacognition. Additionally, one dyad implemented a system whereby they shared metacognitive regulation based on their strengths in learning. Overall, the findings suggested that in order to discover patterns of socially shared metacognition, it is important to investigate metacognition over time. However, it was concluded that more research on socially shared metacognition, from larger data sets, is needed. These findings formed the basis of the second study. In Study II, the specific aim was to investigate whether socially shared metacognition can be reliably identified from a large dataset of collaborative face-to-face mathematical word problem solving sessions by student dyads. We specifically examined different difficulty levels of tasks as well as the function and focus of socially shared metacognition. Furthermore, the presence of observable metacognitive experiences at the beginning of socially shared metacognition was explored. Four dyads participated in the study. Each dyad was comprised of high-achieving 10-year-old students, ranked in the top 11% of their fourth grade peers (n=393). Dyads were from the same data set as in Study I. The dyads worked face-to-face in a computer-supported, game-format learning environment. Problem-solving processes for 251 tasks at three difficulty levels taking place during 56 (30–45 minutes) lessons were video-taped and analysed. Baseline data for this study were 14 675 turns of transcribed verbal and nonverbal behaviours observed in four study dyads. The micro-level analysis illustrated how participants moved between different channels of communication (individual and interpersonal). The unit of analysis was a set of turns, referred to as an ‘episode’. The results indicated that socially shared metacognition and its function and focus, as well as the appearance of metacognitive experiences can be defined in a reliable way from a larger data set by independent coders. A comparison of the different difficulty levels of the problems suggested that in order to trigger socially shared metacognition in small groups, the problems should be more difficult, as opposed to moderately difficult or easy. Although socially shared metacognition was found in collaborative face-to-face problem solving among high-achieving student dyads, more research is needed in different contexts. This consideration created the basis of the research on socially shared metacognition in Studies III and IV. In Study III, the aim was to expand the research on SSMR from face-to-face mathematical problem solving in student dyads to inquiry-based science learning among small groups in an asynchronous computer-supported collaborative learning (CSCL) environment. The specific aims were to investigate SSMR’s evolvement and functions in a CSCL environment and to explore how SSMR emerges at different phases of the inquiry process. Finally, individual student participation in SSMR during the process was studied. An in-depth explanatory case study of one small group of four girls aged 12 years was carried out. The girls attended a class that has an entrance examination and conducts a language-enriched curriculum. The small group solved complex science problems in an asynchronous CSCL environment, participating in research-like processes of inquiry during 22 lessons (á 45–minute). Students’ network discussion were recorded in written notes (N=640) which were used as study data. A set of notes, referred to here as a ‘thread’, was used as the unit of analysis. The inter-coder agreement was regarded as substantial. The results indicated that SSMR emerges in a small group’s asynchronous CSCL inquiry process in the science domain. Hence, the results of Study III were in line with the previous Study I and Study II and revealed that metacognition cannot be reduced to the individual level alone. The findings also confirm that SSMR should be examined as a process, since SSMR can evolve during different phases and that different SSMR threads overlapped and intertwined. Although the classification of SSMR’s functions was applicable in the context of CSCL in a small group, the dominant function was different in the asynchronous CSCL inquiry in the small group in a science activity than in mathematical word problem solving among student dyads (Study II). Further, the use of different analytical methods provided complementary findings about students’ participation in SSMR. The findings suggest that it is not enough to code just a single written note or simply to examine who has the largest number of notes in the SSMR thread but also to examine the connections between the notes. As the findings of the present study are based on an in-depth analysis of a single small group, further cases were examined in Study IV, as well as looking at the SSMR’s focus, which was also studied in a face-to-face context. In Study IV, the general aim was to investigate the emergence of SSMR with a larger data set from an asynchronous CSCL inquiry process in small student groups carrying out science activities. The specific aims were to study the emergence of SSMR in the different phases of the process, students’ participation in SSMR, and the relation of SSMR’s focus to the quality of outcomes, which was not explored in previous studies. The participants were 12-year-old students from the same class as in Study III. Five small groups consisting of four students and one of five students (N=25) were involved in the study. The small groups solved ill-defined science problems in an asynchronous CSCL environment, participating in research-like processes of inquiry over a total period of 22 hours. Written notes (N=4088) detailed the network discussions of the small groups and these constituted the study data. With these notes, SSMR threads were explored. As in Study III, the thread was used as the unit of analysis. In total, 332 notes were classified as forming 41 SSMR threads. Inter-coder agreement was assessed by three coders in the different phases of the analysis and found to be reliable. Multiple methods of analysis were used. Results showed that SSMR emerged in all the asynchronous CSCL inquiry processes in the small groups. However, the findings did not reveal any significantly changing trend in the emergence of SSMR during the process. As a main trend, the number of notes included in SSMR threads differed significantly in different phases of the process and small groups differed from each other. Although student participation was seen as highly dispersed between the students, there were differences between students and small groups. Furthermore, the findings indicated that the amount of SSMR during the process or participation structure did not explain the differences in the quality of outcomes for the groups. Rather, when SSMRs were focused on understanding and procedural matters, it was associated with achieving high quality learning outcomes. In turn, when SSMRs were focused on incidental and procedural matters, it was associated with low level learning outcomes. Hence, the findings imply that the focus of any emerging SSMR is crucial to the quality of the learning outcomes. Moreover, the findings encourage the use of multiple research methods for studying SSMR. In total, the four studies convincingly indicate that a phenomenon of socially shared metacognitive regulation also exists. This means that it was possible to define the concept of SSMR theoretically, to investigate it methodologically and to validate it empirically in two different learning contexts across dyads and small groups. In-depth micro-level case analysis in Studies I and III showed the possibility to capture and analyse in detail SSMR during the collaborative process, while in Studies II and IV, the analysis validated the emergence of SSMR in larger data sets. Hence, validation was tested both between two environments and within the same environments with further cases. As a part of this dissertation, SSMR’s detailed functions and foci were revealed. Moreover, the findings showed the important role of observable metacognitive experiences as the starting point of SSMRs. It was apparent that problems dealt with by the groups should be rather difficult if SSMR is to be made clearly visible. Further, individual students’ participation was found to differ between students and groups. The multiple research methods employed revealed supplementary findings regarding SSMR. Finally, when SSMR was focused on understanding and procedural matters, this was seen to lead to higher quality learning outcomes. Socially shared metacognition regulation should therefore be taken into consideration in students’ collaborative learning at school similarly to how an individual’s metacognition is taken into account in individual learning.