Dans les systèmes cognitifs, le rôle de la mémoire de travail est crucial pour le raisonnement visuel et la prise de décision. D’énormes progrès ont été réalisés dans la compréhension des mécanismes de la mémoire de travail humain/animal, ainsi que dans la formulation de différents cadres de réseaux de neurones artificiels à mémoire augmentée. L’objectif global de notre projet est de former des modèles de réseaux de neurones artificiels capables de consolider la mémoire sur une courte période de temps pour résoudre une tâche de mémoire et les relier à l’activité cérébrale des humains qui ont résolu la même tâche. Le projet est de nature interdisciplinaire en essayant de relier les aspects de l’intelligence artificielle (apprentissage profond) et des neurosciences. La tâche cognitive utilisée est la tâche N-back, très populaire en neurosciences cognitives dans laquelle les sujets sont présentés avec une séquence d’images, dont chacune doit être identifiée pour savoir si elle a déjà été vue ou non. L’ensemble de données d’imagerie fonctionnelle (IRMf) utilisé a été collecté dans le cadre du projet Courtois Neurmod. Nous étudions plusieurs variantes de modèles de réseaux neuronaux récurrents qui apprennent à résoudre la tâche de mémoire de travail N-back en les entraînant avec des séquences d’images. Ces réseaux de neurones entraînés optimisés pour la tâche de mémoire sont finalement utilisés pour générer des représentations de caractéristiques pour les images de stimuli vues par les sujets humains pendant leurs enregistrements tout en résolvant la tâche. Les représentations dérivées de ces réseaux de neurones servent ensuite à créer un modèle de codage pour prédire l’activité IRMf BOLD des sujets. On comprend alors la relation entre le modèle de réseau neuronal et l’activité cérébrale en analysant cette capacité prédictive du modèle dans différentes zones du cerveau impliquées dans la mémoire de travail. Ce travail présente une manière d’utiliser des réseaux de neurones artificiels pour modéliser le comportement et le traitement de l’information de la mémoire de travail du cerveau et d’utiliser les données d’imagerie cérébrale capturées sur des sujets humains lors de la tâche N-back pour potentiellement comprendre certains mécanismes de mémoire du cerveau en relation avec ces modèles de réseaux de neurones artificiels. / In cognitive systems, the role of working memory is crucial for visual reasoning and decision making. Tremendous progress has been made in understanding the mechanisms of the human/animal working memory, as well as in formulating different frameworks of memory augmented artificial neural networks. The overall objective of our project is to train artificial neural network models that are capable of consolidating memory over a short period of time to solve a memory task and relate them to the brain activity of humans who solved the same task. The project is of interdisciplinary nature in trying to bridge aspects of Artificial Intelligence (deep learning) and Neuroscience. The cognitive task used is the N-back task, a very popular one in Cognitive Neuroscience in which the subjects are presented with a sequence of images, each of which needs to be identified as to whether it was already seen or not. The functional imaging (fMRI) dataset used has been collected as a part of the Courtois Neurmod Project. We study multiple variants of recurrent neural network models that learn to remember input images across timesteps. These trained neural networks optimized for the memory task are ultimately used to generate feature representations for the stimuli images seen by the human subjects during their recordings while solving the task. The representations derived from these neural networks are then to create an encoding model to predict the fMRI BOLD activity of the subjects. We then understand the relationship between the neural network model and brain activity by analyzing this predictive ability of the model in different areas of the brain that are involved in working memory. This work presents a way of using artificial neural networks to model the behavior and information processing of the working memory of the brain and to use brain imaging data captured from human subjects during the N-back task to potentially understand some memory mechanisms of the brain in relation to these artificial neural network models.

<p>Advances in Artificial Intelligence (AI), or more precisely on Neural Networks (NNs), and fast processing technologies (e.g. Graphic Processing Units or GPUs) in recent years have positioned NNs as one of the main machine learning algorithms used to solved a diversity of problems in both academia and the industry. While they have been proved to be effective in solving many tasks, the lack of security guarantees and understanding of their internal processing disrupts their wide adoption in general and cybersecurity-related applications. In this dissertation, we present the findings of a comprehensive study aimed to enable the absorption of state-of-the-art NN algorithms in the development of enterprise solutions. Specifically, this dissertation focuses on (1) the development of defensive mechanisms to protect NNs against adversarial attacks and (2) application of NN models for anomaly detection in enterprise networks.</p><p>In this state of affairs, this work makes the following contributions. First, we performed a thorough study of the different adversarial attacks against NNs. We concentrate on the attacks referred to as trojan attacks and introduce a novel model hardening method that removes any trojan (i.e. misbehavior) inserted to the NN models at training time. We carefully evaluate our method and establish the correct metrics to test the efficiency of defensive methods against these types of attacks: (1) accuracy with benign data, (2) attack success rate, and (3) accuracy with adversarial data. Prior work evaluates their solutions using the first two metrics only, which do not suffice to guarantee robustness against untargeted attacks. Our method is compared with the state-of-the-art. The obtained results show our method outperforms it. Second, we proposed a novel approach to detect anomalies using LSTM-based models. Our method analyzes at runtime the event sequences generated by the Endpoint Detection and Response (EDR) system of a renowned security company running and efficiently detects uncommon patterns. The new detecting method is compared with the EDR system. The results show that our method achieves a higher detection rate. Finally, we present a Moving Target Defense technique that smartly reacts upon the detection of anomalies so as to also mitigate the detected attacks. The technique efficiently replaces the entire stack of virtual nodes, making ongoing attacks in the system ineffective.</p><p> </p>

Decentralized Machine Learning could address some problematic facets with Federated Learning. There is no central server acting as an arbiter of whom or what may benefit from Machine Learning models created by the vast amount of data becoming available in recent years. It could also increase the reliability and scalability of Machine Learning systems thereby drawing the benefit of having more data accessible. Gossip Learning is such a protocol, but has primarily been designed with linear models in mind. How does Gossip Learning perform when training Deep Neural Networks? Could it be a viable alternative to Federated Learning? In this thesis, we implement Gossip Learning using two different model merging strategies. We also design and implement two extensions to this protocol with the goal of achieving higher performance when training under churn. The training methods are compared on two tasks: image classification on the Federated Extended MNIST dataset and time- series forecasting on the NN5 dataset. Additionally, we also run an experiment where learners churn, alternating between being available and unavailable. We find that Gossip Learning performs slightly better in settings where learners do not churn but is vastly outperformed in the setting where they do. / Decentraliserad Maskinginlärning kan lösa några problematiska aspekter med Federated Learning. Det finns ingen central server som agerar som domare för vilka som får gagna av Maskininlärningsmodellerna skapad av den stora mäng data som blivit tillgänglig på senare år. Det skulle också kunna öka pålitligheten och skalbarheten av Maskininlärningssystem och därav dra nytta av att mer data är tillgänglig. Gossip Learning är ett sånt protokoll, men det är primärt designat med linjära modeller i åtanke. Hur presterar Gossip Learning när man tränar Djupa Neurala Nätverk? Kan det vara ett möjligt alternativ till Federated Learning? I det här exjobbet implementerar vi Gossip Learning med två olika modelsammanslagningstekniker. Vi designar och implementerar även två tillägg till protokollet med målet att uppnå bättre prestanda när man tränar i system där noder går ner och kommer up. Träningsmetoderna jämförs på två uppgifter: bildklassificering på Federated Extended MNIST datauppsättningen och tidsserieprognostisering på NN5 datauppsättningen. Dessutom har vi även experiment då noder alternerar mellan att vara tillgängliga och otillgängliga. Vi finner att Gossip Learning presterar marginellt bättre i miljöer då noder alltid är tillgängliga men är kraftigt överträffade i miljöer då noder alternerar mellan att vara tillgängliga och otillgängliga.

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.

Fraud is a common crime within the insurance industry, and insurance companies want to quickly identify fraudulent claimants as they often result in higher premiums for honest customers. Due to the digital transformation where the sheer volume and complexity of available data has grown, manual fraud detection is no longer suitable. This work aims to automate the detection of fraudulent claimants and gain practical insights into fraudulent behavior using unsupervised anomaly detection, which, compared to supervised methods, allows for a more cost-efficient and practical application in the insurance industry. To obtain interpretable results and benefit from the temporal dependencies in human behavior, we propose two variations of LSTM based autoencoders to classify sequences of insurance claims. Autoencoders can provide feature importances that give insight into the models' predictions, which is essential when models are put to practice. This approach relies on the assumption that outliers in the data are fraudulent. The models were trained and evaluated on a dataset we engineered using data from a Swedish insurance company, where the few labeled frauds that existed were solely used for validation and testing. Experimental results show state-of-the-art performance, and further evaluation shows that the combination of autoencoders and LSTMs are efficient but have similar performance to the employed baselines. This thesis provides an entry point for interested practitioners to learn key aspects of anomaly detection within fraud detection by thoroughly discussing the subject at hand and the details of our work. / <p>Gjordes digitalt via Zoom. </p>

Using AI to automatically describe images is a challenging task. The aim of this study has been to compare the use of character-based language models with one of the current state-of-the-art token-based language models, im2txt, to generate image captions, with focus on morphological correctness. Previous work has shown that character-based language models are able to outperform token-based language models in morphologically rich languages. Other studies show that simple multi-layered LSTM-blocks are able to learn to replicate the syntax of its training data. To study the usability of character-based language models an alternative model based on TensorFlow im2txt has been created. The model changes the token-generation architecture into handling character-sized tokens instead of word-sized tokens. The results suggest that a character-based language model could outperform the current token-based language models, although due to time and computing power constraints this study fails to draw a clear conclusion. A problem with one of the methods, subsampling, is discussed. When using the original method on character-sized tokens this method removes characters (including special characters) instead of full words. To solve this issue, a two-phase approach is suggested, where training data first is separated into word-sized tokens where subsampling is performed. The remaining tokens are then separated into character-sized tokens. Future work where the modified subsampling and fine-tuning of the hyperparameters are performed is suggested to gain a clearer conclusion of the performance of character-based language models.

No description available.

Developing and improving upon a good empirical model for an engine can be time-consuming and costly. The goal of this thesis has been to evaluate data-driven modelling, specifically neural networks, to see how well it can handle training for some static models like the mass flow of air into the cylinder, mean effective pressure and pump mean effective pressure but also for transient modelling, specifically the exhaust gas temperature. These models are evaluated against the classical empirical models to see if neural networks are a viable modelling option. This is done with five different types of neural networks which are trained. These are the feed-forward neural network, Nonlinear autoregressive exogenous model network, layer recurrent network, long short term memory network and gated recurrent network.The inputs were determined by looking at more simple physical models but also looking at the covariance to determine the usefulness of the input. If the calculation time is small for the specific network, the neural network structure is tested and optimized by training many networks and finding the median/mean result for that specific test.The result has shown that the static models are handled very well by the most simple feed-forward network. For the exhaust temperature, both NARX and Layer recurrent network could predict and handle it well giving results very close to the empirical models and could be a viable option for transient modelling, on the other hand, Long short term memory, gated recurrent network and the feed-forward network had trouble predicting the exhaust gas temperature and returned bad results while training.

Nowadays, a significant proportion of the population in the world is affected by cerebral diseases like epilepsy. In this study, frequency domain features of electroencephalography (EEG) signals were studied and analyzed, with a view being able to detect epileptic seizures more easily. The power spectrum and spectrogram were determined by using fast fourier transform (FFT) and the scalogram was found by performing continuous wavelet transform (CWT) on the testing EEG signal. In addition, two schemes, i.e. method 1 and method 2, were implemented for detecting epileptic seizures and the applicability of the two methods to electrocardiogram (ECG) signals were tested. A third method for anomaly detection in ECG signals was tested. / En signifikant del av population påverkas idag av neurala sjukdomar som epilepsi. I denna studie studerades och analyserades egenskaper inom frekvensdomänen av elektroencefalografi (EEG), med sikte på att lättare kunna upptäcka epileptiska anfall. Effektspektrumet och spektrogramet bestämdes med hjälp av en snabb fouriertransform och skalogrammet hittades genom att genomföra en kontinuerlig wavelet transform (CWT) på testsignalen från EEGsignalen. I addition till detta skapades två system, metod 1 och metod 2, som implementerades för att upptäcka epileptiska anfall. Användbarheten av dessa två metoder inom elektrokardiogramsignaler (ECG) testades. En tredje metod för anomalidetektering i ECGsignaler testades.

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.