No description available.

No description available.

The hypothesis that similar chemicals exert similar biological activities has been widely adopted in the field of drug discovery and development. Quantitative Structure-Activity Relationship (QSAR) models have been used ubiquitously in drug discovery to understand the function of chemicals in biological systems. A common QSAR modeling method calculates similarity scores between chemicals to assess their biological function. However, due to the fact that some chemicals can be similar and yet have different biological activities, or conversely can be structurally different yet have similar biological functions, various methods have instead been developed to quantify chemical similarity at the functional level. Quantmap is one such method, which utilizes biological databases to quantify the biological similarity between chemicals. Quantmap uses quantitative molecular network topology analysis to cluster chemical substances based on their bioactivities. This method by itself, unfortunately, cannot assign new chemicals (those which may not yet have biological data) to the derived clusters. Owing to the fact that there is a lack of biological data for many chemicals, deep learning models were explored in this project with respect to their ability to correctly assign unknown chemicals to Quantmap clusters. The deep learning methods explored included both convolutional and recurrent neural networks. Transfer learning/pretraining based approaches and data augmentation methods were also investigated. The best performing model, among those considered, was the Seq2seq model (a recurrent neural network containing two joint networks, a perceiver and an interpreter network) without pretraining, but including data augmentation.

Within the realm of computer vision tasks, Convolutional Neural Networks (CNN) and Transformers represent two predominant methodologies, often subject to extensive comparative analyses elucidating their respective merits and demerits. This thesis embarks on an exploration of these two models within the framework of continual learning, with a specific focus on their propensities for resisting catastrophic forgetting. We hypothesize that Transformer models exhibit a higher resilience to catastrophic forgetting in comparison to their CNN counterparts. To substantiate this hypothesis, a meticulously crafted experimental design was implemented, involving the selection of diverse models and continual learning approaches, and careful tuning of the networks to ensure an equitable comparison. In the majority of conducted experiments, encompassing both the contexts of class incremental learning settings and task incremental learning settings, our results substantiate the aforementioned hypothesis. Nevertheless, the insights garnered also underscore the necessity for more exhaustive and encompassing experimental evaluations to fully validate the asserted hypothesis. / Inom datorseende är Convolutional Neural Networks (CNN) och Transformers två dominerande metoder, som ofta är föremål för omfattande jämförande analyser som belyser deras respektive fördelar och nackdelar. Denna avhandling utforskar dessa två modeller inom ramen för kontinuerligt lärande, med särskilt fokus på deras benägenhet att motstå katastrofal glömska. Vi antar att Transformer-modeller uppvisar en ökad motståndskraft mot katastrofal glömska i jämförelse med deras CNN-motsvarigheter. För att underbygga denna hypotes implementerades en noggrant utformad experimentell design, som involverade val av olika modeller och kontinuerliga inlärningstekniker, och noggrann inställning av nätverken för att säkerställa en rättvis jämförelse. I majoriteten av de genomförda experimenten, som omfattade både inkrementell klassinlärning och inkrementell uppgiftsinlärning, bekräftade våra resultat den ovannämnda hypotesen. De insikter vi fått understryker dock också behovet av mer uttömmande och omfattande experimentella utvärderingar för att fullt ut validera den påstådda hypotesen.

Объектом исследования являются методы машинного обучения, позволяющие фильтровать данные, и методы разработки информационных платформ. Фильтрация данных подобного типа применяется в такой области, как поисковые системы, чтобы на основе запроса выдать пользователю релевантные результаты. Предмет исследования – разработка модели машинного обучения, фильтрующей текстовые данные, и информационной платформы для отображения отфильтрованных данных. Особенностями исследования являются открытая реализация полного проекта, то есть она доступна каждому, и возможность его модификации. Для обучения модели был использован самостоятельно составленный набор научных работ, информационная платформа была разработана с нуля. Итоговая модель LSTM, выбранная методом сравнения метрик, показала результат предсказания соответствия целевой тематике в 90%, что позволяет говорить о ее возможном внедрении в соответствующие Интернет-ресурсы, так как они гарантированно уменьшат объем научных работ, проверяемых вручную. / The object of the research is machine learning methods that allow filtering text data obtained from the information platform. Filtering of this type of data is used in such an area as search engines to give relevant results to the user based on a query. Within the framework of this dissertation, it was proposed to apply machine learning methods to filter a set of scientific papers based on their title and target label in the form of the subject of the work. The features of the study are the open implementation of the full project, that is, it is available to everyone, and the possibility of its modification. A self-compiled set of scientific papers was used to train the model, the information platform was developed from scratch. The final LSTM model, chosen by the method of comparing metrics, showed the result of predicting compliance with the target topic in 95%, which allows us to talk about its possible implementation in the relevant Internet resources, since they are guaranteed to reduce the volume of scientific papers checked manually.

[pt] O desmatamento é uma das principais causas de redução da biodiversidade, mudança climática e outros fenômenos destrutivos. Assim, a detecção antecipada de desmatamento é de suma importância. Técnicas baseadas em imagens de satélite são uma das opções mais iteresantes para esta aplicação. No entanto, muitos trabalhos desenvolvidos incluem algumas operações manuais ou dependência de um limiar para identificar regiões que sofrem desmatamento ou não. Motivado por este cenário, a presente dissertação apresenta uma avaliação de métodos para detecção automática de desmatamento, especificamente de Early Fusion (EF) Convolutional Network, Siamese Convolutional Network (SN), Convolutional Support Vector Machine (CSVM) e Support Vector Machine (SVM), o último tomado como baseline. Todos os métodos foram avaliados em regiões dos biomas brasileiros Amazônia e Cerrado. Duas imagens Landsat 8 adquiridas em diferentes datas foram utilizadas nos experimentos, e também o impacto do tamanho do conjunto de treinamento foi analisado. Os resultados demonstraram que as abordagens baseadas no Deep Learning superaram claramente o baseline SVM em termos de pontuação F1-score e Overrall Accuracy, com uma superioridade de SN e EF sobre CSVM e SVM. Da mesma forma, uma redução do efeito sal e pimenta nos mapas de mudança gerados foi notada devido, principalmente ao aumento de amostras nos conjuntos de treinamento. Finalmente, realizou-se uma análise visando avaliar como os métodos podem reduzir o esforço humano na inspeção visual das áreas desmatadas. / [en] Deforestation is one of the main causes of biodiversity reduction, climate change, among other destructive phenomena. Thus, early detection of deforestation processes is of paramount importance. Techniques based on satellite images are one of the most attractive options for this application. However, many works developed include some manual operations or dependency on a threshold to identify regions that suffer deforestation or not. Motivated by this scenario, the present dissertation presents an evaluation of methods for automatic deforestation detection, specifically Early Fusion (EF) Convolutional Network, Siamese Convolutional Network (SN), Convolutional Support Vector Machine (CSVM) and Support Vector Machine (SVM), taken as the baseline. These methods were evaluated in regions of Brazilian Amazon and Cerrado Biomes. Two Landsat 8 images acquired at different dates were used in the experiments, and the impact of training set size was also analyzed. The results demonstrated that Deep Learning-based approaches clearly outperformed the SVM baseline in our approaches, both in terms of F1-score and Overall Accuracy, with the superiority of SN and EF over CSVM and SVM. In the same way, a reduction of the salt-and-pepper effect in the generated probabilistic change maps was noticed due, mainly, to the increase of samples in the training sets. Finally, an analysis was carried out to assess how the methods can reduce the time invested in the visual inspection of deforested areas.

CSAW-M is a dataset that contains about 10 000 x-ray images created from mammograms. Mammograms are used to identify patients with breast cancer through a screening process with the goal of catching cancer tumours early. Modern convolutional neural networks are very sophisticated and capable of identifying patterns nearly indistinguishable to humans. CSAW-M doesn’t contain images of active cancer tumours, rather, whether the patient will develop cancer or not. Classification tasks such as this are known to require large datasets for training, which is cumbersome to acquire in the biomedical domain. In this paper we investigate how classification performance of non-trivial classification tasks scale with the size of available annotated images. To research this, a wide range of data-sets are generated from CSAW-M, with varying sample size and cancer types. Three different convolutional neural networks were trained on all data-sets. The study showed that classification performance does increase with the size of the annotated dataset. All three networks generally improved their prediction on the supplied benchmarking dataset. However, the improvements were very small and the research question could not be conclusively answered. The primary reasons for this was the challenging nature of the classification task, and the size of the data-set. Further research is required to gain more understanding of how much data is needed to yield a usable model. / CSAW-M är ett dataset som innehåller ungefär 10 000 röntgenbilder skapade från ett stort antal mammografier. Mammografi används för att identifiera patienter med bröstcancer genom en screeningprocess med målet att fånga cancerfall tidigt. Moderna konvolutionella neurala nätverk är mycket sofistikerade och kan tränas till att identifiera mönster i bilder mycket bättre än människor. CSAW-M innehåller inga bilder av cancertumörer, utan istället data på huruvida patienten kommer att utveckla cancer eller inte. Klassificeringsuppgifter som denna är kända för att kräva stora datamängder för träning, vilket är svårt att införskaffa inom den biomedicinska domänen. I denna artikel undersöker vi hur klassificerings prestanda för svåra klassificeringsuppgifter skalar med storleken på tillgänglig annoterad data. För att undersöka detta, genererades ett antal nya dataset från CSAW-M, med varierande storleksurval och cancertyp. Tre olika konvolutionella neurala nätverk tränades på alla nya data-set. Studien visar att klassificeringsprestanda ökar med storleken på den annoterade datamängden. Alla tre nätverk förbättrade generellt sin klassificeringsprestanda desto större urval som gjordes från CSAW-M. Förbättringarna var dock små och den studerade frågan kunde inte besvaras fullständigt. De främsta anledningarna till detta var klassificeringsuppgiftens utmanande karaktär och storleken på det tillgängliga datat i CSAW-M. Ytterligare forskning krävs för att få mer förståelse för hur mycket data som behövs för att skapa en användbar modell.

Evidence-based software engineering (EBSE) solutions for drug discovery that are effective, affordable, and accessible all-in-one are lacking. This thesis chronicles the progression and accomplishments of the AiDA (Artificially-intelligent Desktop Assistant) functional artificial intelligence (AI) project for the purposes of drug discovery in the challenging acute myeloid leukemia context (AML). AiDA is a highly automated combined natural language processing (NLP) and spreadsheet feature extraction solution that harbours potential to disrupt the state of current research investigation methods using big data and aggregated literature. The completed work includes a text-to-function (T2F) NLP method for automated text interpretation, a ranked-list algorithm for multi-dataset analysis, and a custom multi-purpose neural network engine presented to the user using an open-source graphics engine. Validation of the deep learning engine using MNIST and CIFAR machine learning benchmark datasets showed performance comparable to state-of-the-art libraries using similar architectures. An n-dimensional word embedding method for the handling of unstructured natural language data was devised to feed convolutional neural network (CNN) models that over 25 random permutations correctly predicted functional responses to up to 86.64% of over 300 validation transcripts. The same CNN NLP infrastructure was then used to automate biomedical context recognition in >20000 literature abstracts with up to 95.7% test accuracy over several permutations. The AiDA platform was used to compile a bidirectional ranked list of potential gene targets for pharmaceuticals by extracting features from leukemia microarray data, followed by mining of the PubMed biomedical citation database to extract recyclable pharmaceutical candidates. Downstream analysis of the candidate therapeutic targets revealed enrichments in AML- and leukemic stem cell (LSC)-related pathways. The applicability of the AiDA algorithms in whole and part to the larger biomedical research field is explored. / Thesis / Master of Science (MSc) / Lead generation is an integral requirement of any research organization in all fields and is typically a time-consuming and therefore expensive task. This is due to the requirement of human intuition to be applied iteratively over a large body of evidence. In this thesis, a new technology called the Artificially-intelligent Desktop Assistant (AiDA) is explored in order to provide a large number of leads from accumulated biomedical information. AiDA was created using a combination of classical statistics, deep learning methods, and modern graphical interface engineering. It aims to simplify the interface between the researcher and an assortment of bioinformatics tasks by organically interpreting written text messages and responding with the appropriate task. AiDA was able to identify several potential targets for new pharmaceuticals in acute myeloid leukemia (AML), a cancer of the blood, by reading whole-genome data. It then discovered appropriate therapeutics by automatically scanning through the accumulated body of biomedical research papers. Analysis of the discovered drug targets shows that together, they are involved in key biological processes that are known by the scientific community to be involved in leukemia and other cancers.

New satellite missions have provided High Resolution (HR) Satellite Image Time Series (SITS), offering detailed spatial, spectral, and temporal information for effective monitoring of diverse Earth features including weather, landforms, oceans, vegetation, and agricultural practices. SITS can be used for an accurate understanding of the Land Cover (LC) behavior and providing the possibility of precise mapping of LCs. Moreover, HR SITS presents an unprecedented possibility for the creation and modification of HR Land Cover Change (LCC) and Land Cover Transition (LCT) maps. For the long-term scale, spanning multiple years, it becomes feasible to analyze LCC and the LCTs occurring between consecutive years. Existing methods in literature often analyze bi-temporal images and miss the valuable multi-temporal/multi-annual information of SITS that is crucial for an accurate SITS analysis. As a result, HR SITS necessitates a paradigm shift in processing and methodology development, introducing new challenges in data handling. Yet, the creation of techniques that can effectively manage the high spatial correlation and complementary temporal resolutions of pixels remains paramount. Moreover, the temporal availability of HR data across historical and current archives varies significantly, creating the need for an effective preprocessing to account for factors like atmospheric and radiometric conditions that can affect image reflectance and their applicability in SITS analysis. Flexible and automatic SITS analysis methods can be developed by paying special attention to handling big amounts of data and modeling the correlation and characterization of SITS in space and time. Novel methods should deal with data preparation and pre-processing at large-scale from end-to-end by introducing a set of steps that guarantee reliable SITS analysis while upholding the computational efficiency for a feasible SITS analysis. In this context, the recent strides in deep learning-based frameworks have demonstrated their potential across various image processing tasks, and thus the high relevance for addressing SITS analysis. Deep learning-based methods can be supervised or unsupervised considering their learning process. Supervised deep learning methods rely on labeled training data, which can be impractical for large-scale multi-temporal datasets, due to the challenges of manual labeling. In contrast, unsupervised deep learning methods are favored as they can automatically discover temporal patterns and changes without the need for labeled samples, thereby reducing the computational load, making them more suitable for handling extensive SITS. In this scenario, the objectives of this thesis are mainly three. Firstly, it seeks to establish a robust and reliable framework for the precise mapping of LCs by designing novel techniques for time series analysis. Secondly, it aims to utilize the capacities of unsupervised deep learning methods, such as pretrained Convolutional Neural Networks (CNNs), to construct a comprehensive methodology for Change Detection (CD), thereby mitigating complexity and reducing computational requirements in comparison with supervised methods. This involves the efficient extraction of spatial, spectral, and temporal features from complex multi-temporal, multi-spectral SITS. Lastly, the thesis endeavors to develop novel methods for analyzing LCCs occurring over extended time periods, spanning multiple years. This multifaceted approach encompasses the detection of changes, timing identification, and classification of the specific types of LCTs. The efficacy of the innovative methodologies and associated techniques is showcased through a series of experiments conducted on HR SITS datasets, including those from Sentinel-2 and Landsat. These experiments reveal significant enhancements when compared to existing methods that represent the current state-of-the-art.

We investigate the novel idea of using computer vision to predict future stock price movement, which is performed by training a convolutional neural network (CNN) to detect patterns in images of stock graphs. Subsequently, we create a portfolio strategy based on the CNN stock price predictions to see if these predictions can generate alpha for investors. We apply this method in the Swedish stock market and evaluate the performance of CNN portfolios across two different exchanges and various stock indices segmented by market capitalisation. Our findings show that trading based on CNN predictions can outperform our benchmarks and generate positive alpha. Most of our portfolios generate positive alpha before transaction costs, while one also generates positive alpha after deducting transaction costs. Further, our results demonstrate that CNN models are capable of successfully generalising their trained knowledge, being able to detect information in stock graphs it has never seen before. This suggests that CNN models are not limited to the characteristics present in their training data, indicating that models trained under one set of market conditions can also be effective in a different market scenario. Our resultsfurther strengthen the overall findings of other researchers utilising similar methods as ours.