Browsing by Subject "deep learning"

The detection heuristic in contemporary machine learning Windows malware classifiers is typically based on the static properties of the sample. In contrast, simultaneous utilization of static and behavioral telemetry is vaguely explored. We propose a hybrid model that employs dynamic malware analysis techniques, contextual information as an executable filesystem path on the system, and static representations used in modern state-of-the-art detectors. It does not require an operating system virtualization platform. Instead, it relies on kernel emulation for dynamic analysis. Our model reports enhanced detection heuristic and identify malicious samples, even if none of the separate models express high confidence in categorizing the file as malevolent. For instance, given the $0.05\%$ false positive rate, individual static, dynamic, and contextual model detection rates are $18.04\%$, $37.20\%$, and $15.66\%$. However, we show that composite processing of all three achieves a detection rate of $96.54\%$, above the cumulative performance of individual components. Moreover, simultaneous use of distinct malware analysis techniques address independent unit weaknesses, minimizing false positives and increasing adversarial robustness. Our experiments show a decrease in contemporary adversarial attack evasion rates from $26.06\%$ to $0.35\%$ when behavioral and contextual representations of sample are employed in detection heuristic.

Automatic headline generation has the potential to significantly assist editors charged with head- lining articles. Approaches to automation in the headlining process can range from tools as creative aids, to complete end to end automation. The latter is difficult to achieve as journalistic require- ments imposed on headlines must be met with little room for error, with the requirements depending on the news brand in question. This thesis investigates automatic headline generation in the context of the Finnish newsroom. The primary question I seek to answer is how well the current state of text generation using deep neural language models can be applied to the headlining process in Finnish news media. To answer this, I have implemented and pre-trained a Finnish generative language model based on the Transformer architecture. I have fine-tuned this language model for headline generation as autoregression of headlines conditioned on the article text. I have designed and implemented a variation of the Diverse Beam Search algorithm, with additional parameters, to perform the headline generation in order to generate a diverse set of headlines for a given text. The evaluation of the generative capabilities of this system was done with real world usage in mind. I asked domain-experts in headlining to evaluate a generated set of text-headline pairs. The task was to accept or reject the individual headlines in key criteria. The responses of this survey were then quantitatively and qualitatively analyzed. Based on the analysis and feedback, this model can already be useful as a creative aid in the newsroom despite being far from ready for automation. I have identified concrete improvement directions based on the most common types of errors, and this provides interesting future work.

A recent machine learning technique called federated learning (Konecny, McMahan, et. al., 2016) offers a new paradigm for distributed learning. It consists of performing machine learning on multiple edge devices and simultaneously optimizing a global model for all of them, without transmitting user data. The goal for this thesis was to prove the benefits of applying federated learning to forecasting telecom key performance indicator (KPI) values from radio network cells. After performing experiments with different data sources' aggregations and comparing against a centralized learning model, the results revealed that a federated model can shorten the training time for modelling new radio cells. Moreover, the amount of transferred data to a central server is minimized drastically while keeping equivalent performance to a traditional centralized model. These experiments were performed with multi-layer perceptron as model architecture after comparing its performance against LSTM. Both, input and output data were sequences of KPI values.

Combining data from visual and inertial sensors effectively reduces inherent errors in each modality, enhancing the robustness of sensor-fusion for accurate 6-DoF motion estimation over extended periods. While traditional SfM and SLAM frameworks are well established in literature and real-world applications, purely end-to-end learnable SfM and SLAM networks are still scarce. The adaptability of fully trained models in system configuration and navigation setup holds great potential for future developments in this field. This thesis introduces and assesses two novel end-to-end trainable sensor-fusion models using a supervised learning approach, tested on established navigation benchmarks and custom datasets. The first model utilizes optical flow, revealing its limitations in handling complex camera movements present in pedestrian motion. The second model addresses these shortcomings by using feature point-matching and a completely original design.

Atomic force microscopy (AFM) is a widely utilized characterization method capable of capturing atomic level detail in individual organic molecules. However, an AFM image contains relatively little information about the deeper atoms in a molecule and thus interpretation of AFM images of non-planar molecules offers significant challenges for human experts. An end-to-end solution starting from an AFM imaging system ending in an automated image interpreter would be a valuable asset for all research utilizing AFM. Machine learning has become a ubiquitous tool in all areas of science. Artificial neural networks (ANNs), a specific machine learning tool, have also arisen as a popular method many fields including medical imaging, self-driving cars and facial recognition systems. In recent years, progress towards interpreting AFM images from more complicated samples has been made utilizing ANNs. In this thesis, we aim to predict sample structures from AFM images by modeling the molecule as a graph and using a generative model to build the molecular structure atom-by-atom and bond-by-bond. The generative model uses two types of ANNs, a convolutional attention mechanism to process the AFM images and a graph neural network to process the generated molecule. The model is trained and tested using simulated AFM images. The results of the thesis show that the model has the capability to learn even slight details from complicated AFM images, especially when the model only adds a single atom to the molecule. However, there are challenges to overcome in the generative model for it to become a part of a fully capable end-to-end AFM process.

Independent Component Analysis (ICA) aims to separate the observed signals into their underlying independent components responsible for generating the observations. Most research in ICA has focused on continuous signals, while the methodology for binary and discrete signals is less developed. Yet, binary observations are equally present in various fields and applications, such as causal discovery, signal processing, and bioinformatics. In the last decade, Boolean OR and XOR mixtures have been shown to be identifiable by ICA, but such models suffer from limited expressivity, calling for new methods to solve the problem. In this thesis, "Independent Component Analysis for Binary Data", we estimate the mixing matrix of ICA from binary observations and an additionally observed auxiliary variable by employing a linear model inspired by the Identifiable Variational Autoencoder (iVAE), which exploits the non-stationarity of the data. The model is optimized with a gradient-based algorithm that uses second-order optimization with limited memory, resulting in a training time in the order of seconds for the particular study cases. We investigate which conditions can lead to the reconstruction of the mixing matrix, concluding that the method is able to identify the mixing matrix when the number of observed variables is greater than the number of sources. In such cases, the linear binary iVAE can reconstruct the mixing matrix up to order and scale indeterminacies, which are considered in the evaluation with the Mean Cosine Similarity Score. Furthermore, the model can reconstruct the mixing matrix even under a limited sample size. Therefore, this work demonstrates the potential for applications in real-world data and also offers a possibility to study and formalize identifiability in future work. In summary, the most important contributions of this thesis are the empirical study of the conditions that enable the mixing matrix reconstruction using the binary iVAE, and the empirical results on the performance and efficiency of the model. The latter was achieved through a new combination of existing methods, including modifications and simplifications of a linear binary iVAE model and the optimization of such a model under limited computational resources.

Visual fashion understanding (VFU) is a discipline which aims to solve tasks related to clothing recognition, such as garment categorization, garment’s attributes prediction or clothes retrieval, with the use of computer vision algorithms trained on fashion-related data. Having surveyed VFU- related scientific literature, I conclude that, because of the fact that at the heart of all VFU tasks is the same issue of visually understanding garments, those VFU tasks are in fact related. I present a hypothesis that building larger multi-task learning models dedicated to predicting multiple VFU tasks at once might lead to better generalization properties of VFU models. I assess the validity of my hypothesis by implementing two deep learning solutions dedicated primarily to category and attribute prediction. First solution uses multi-task learning concept of sharing features from ad- ditional branch dedicated to localization task of landmarks’ position prediction. Second solution does not share knowledge from localization branch. Comparison of those two implementations con- firmed my hypothesis, as sharing knowledge between tasks increased category prediction accuracy by 53% and attributes prediction recall by 149%. I conclude that multi-task learning improves generalization properties of deep learning-based visual fashion understanding models across tasks.

Mammography is used as an early detection system for breast cancer, which is one of the most common types of cancer, regardless of one’s sex. Mammography uses specialised X-ray machines to look into the breast tissue for possible tumours. Due to the machine’s set-up as well as to reduce the radiation patients are exposed to, the number of X-ray measurements collected is very restricted. Reconstructing the tissue from this limited information is referred to as limited angle tomography. This is a complex mathematical problem and ordinarily leads to poor reconstruction results. The aim of this work is to investigate how well a neural network whose structure utilizes pre-existing models and known geometry of the problem performs at this task. In this preliminary work, we demonstrate the results on simulated two-dimensional phantoms and discuss the extension of the results to 3-dimensional patient data.

As privacy gains consensus in the field of machine learning, numerous algorithms, such as differentially private stochastic gradient descent (DPSGD), have emerged to ensure privacy guarantees. Concurrently, fairness is garnering increasing attention, prompting research aimed at achieving fairness within the constraints of differential privacy. This thesis delves into algorithms designed to enhance fairness in the realm of differentially private deep learning and explores their mechanisms. It examines the role of normalization, a technique applied to these algorithms in practice, to elucidate its impact on fairness. Additionally, this thesis formalizes a hyperparameter tuning protocol to accurately assess the performance of these algorithms. Experiments across various datasets and neural network architectures were conducted to test our hypotheses under this tuning protocol. The decoupling of hyperparameters, allowing each to independently control specific properties of the algorithm, has proven to enhance performance. However, certain mechanisms, such as discarding samples with large norms and allowing unbounded hyperparameter adaptation, may significantly compromise fairness. Our experiments also confirm the critical role of hyperparameter values in influencing fairness, emphasizing the necessity of precise tuning to ensure equitable outcomes. Additionally, we observed differential convergence rates across algorithms, which affect the number of trials needed to identify optimal hyperparameter settings. This thesis aims to offer detailed perspectives on understanding fairness in differentially private deep learning and provides insights into designing algorithms that can more effectively enhance fairness.

Foundation models have the potential to reduce the level of supervision required for medical image segmentation tasks. Currently, the medical image segmentation field still largely relies on supervised, task specific models. The aim of this thesis is to investigate if a foundation model, the Segment Anything Model (SAM), can be used to reduce the level of supervision needed for medical image segmentation. The main goal of this thesis is to see if the annotation workload required to generate labeled medical segmentation datasets can be significantly reduced with the help of Segment Anything Model. The second goal of this thesis is to validate the zero-shot performance of the Segment Anything Model on a medical segmentation dataset. A UNet model is used as a baseline. The results of this thesis give positive feedback on SAM's ability to be used as a tool for medical image annotation. During the experiments, it was found that especially for homogeneous, clearly outlined tasks, like organs, using ''pseudo labels'' generated by SAM for training a UNet model resulted in comparable accuracy with training a UNet model on human-annotated labels. Furthermore, the results show that zero-shot SAM has somewhat comparable performance to UNet, and even beats UNet in two of the experimented tasks. For one complexly structured task, SAM and UNet with pseudo labels, trained using SAM's masks, fail to produce accurate results. It is notable that some of the tasks have small training dataset sizes, which limits the test accuracy of UNet. The results are in accordance with recent literature which shows that zero-shot SAM can have comparable performance to state-of-the-art models with large and distinct objects, but when it comes to small, complex structures, SAM is not up to par accuracy-wise to the state-of-the-art medical segmentation models.