Gradient-based adversarial attacks on neural networks threaten extremely critical applications such as medical diagnosis and biometric authentication. These attacks use the gradient of the neural network to craft imperceptible perturbations to be added to the test data, in an attempt to decrease the accuracy of the network. We propose a defense to combat such attacks, which can be modified to reduce the training time of the network by as much as 71%, and can be further modified to reduce the training time of the defense by as much as 19%. Further, we address the threat of uncertain behavior on the part of the attacker, a threat previously overlooked in the literature that considers mostly white box scenarios. To combat uncertainty on the attacker's part, we train our defense with an ensemble of attacks, each generated with a different attack algorithm, and using gradients of distinct architecture types. Finally, we discuss how we can prevent the attacker from breaking the defense by estimating the gradient of the defense transformation.

Detta arbete fokuserar på prediktion av cykeltrafik under en månad på en given gata i Malmö med hjälp av maskininlärning. Algoritmen som används är Python-implementering av Stödvektormaskin (Support Vector Machine) från Scikit . Data som används är antalet cyklister/dag under 2006-2013 från en cykel-barometer som är placerad på Kaptensgatan i Malmö. Barometerns funktion är att räkna antalet cyklar som passerar samt registrera tiden. I vår studie undersöker vi hur precision av prediktionen av antalet cyklister varje dag under fyra veckor i oktober 2013, mätt med metoderna RMSE och MAPE, beror av valet av indata (cykeldata och angivelse av veckodag). Ett antal experiment med olika kombinationer av indata och representanter av veckodagar genomfördes. Resultaten visar att testet med störst indata-mängd och veckodagar, angivet som 1-7, gav bäst prediktion. / This work focuses on the prediction of bicycle traffic for a month on a given street in Malmö by means of machine learning. The algorithm used is the Python implementation of Support Vector Machine from Scikit. The data used is the number of cyclists / day during 2006-2013 from a cycle barometer placed on Kaptensgatan in Malmö. The function of the barometer is to count the number of cycles that pass and register the time. In our study we investigate how precision of the prediction of the number of cyclists each day for four weeks in October 2013, measured by the RMSE and MAPE methods, depends on the choice of input data (cycle data and the weekday indication). A number of experiments with different combinations of input data and representatives of weekdays were conducted. The results show that the test with the largest input amount and week days indicated as 1-7 gave the best prediction.

We study online active learning under the Bernstein condition for bounded general losses and offer a solution for online variance estimation. Our suggested algorithm is based on IWAL (Importance Weighted Active Learning) which utilizes the online variance estimation technique to shrink the hypothesis set. For our algorithm, we provide a fallback guarantee and prove that in the case that R(f*) is small, it will converge faster than passive learning, where R(f*) is the risk of the best hypothesis in the hypothesis class. Finally, in the special case of zero-one loss exponential improvement is achieved in label complexity over passive learning. / Graduate

Deep Neural Networks (DNNs) have become ubiquitous due to their performance on prediction and classification problems. However, they face a variety of threats as their usage spreads. Model extraction attacks, which steal DNN models, endanger intellectual property, data privacy, and security. Previous research has shown that system-level side channels can be used to leak the architecture of a victim DNN, exacerbating these risks. We propose a novel DNN architecture extraction attack, called EZClone, which uses aggregate rather than time-series GPU profiles as a side-channel to predict DNN architecture. This approach is not only simpler, but also requires less adversary capability than earlier works. We investigate the effectiveness of EZClone under various scenarios including reduction of attack complexity, against pruned models, and across GPUs with varied resources. We find that EZClone correctly predicts DNN architectures for the entire set of PyTorch vision architectures with 100\% accuracy. No other work has shown this degree of architecture prediction accuracy with the same adversarial constraints or using aggregate side-channel information. Prior work has shown that, once a DNN has been successfully cloned, further attacks such as model evasion or model inversion can be accelerated significantly. Then, we evaluate several mitigation techniques against EZClone, showing that carefully inserted dummy computation reduces the success rate of the attack.

Metagenomics is the study of the totality of the complete genetic elements discovered from a defined environment. Different from traditional microbiology study, which only analyzes a small percent of microbes that could survive in laboratory, metagenomics allows researchers to get entire genetic information from all the samples in the communities. So metagenomics enables understanding of the target environments and the hidden relationships between bacteria and diseases. In order to efficiently analyze the metagenomics data, cutting-edge technologies for analyzing the relationships among microbes and communities are required. To overcome the challenges brought by rapid growth in metagenomics datasets, advances in novel methodologies for interpreting metagenomics data are clearly needed. The first two chapters of this dissertation summarize and compare the widely-used methods in metagenomics and integrate these methods into pipelines. Properly analyzing metagenomics data requires a variety of bioinformatcis and statistical approaches to deal with different situations. The raw reads from sequencing centers need to be processed and denoised by several steps and then be further interpreted by ecological and statistical analysis. So understanding these algorithms and combining different approaches could potentially reduce the influence of noises and biases at different steps. And an efficient and accurate pipeline is important to robustly decipher the differences and functionality of bacteria in communities. Traditional statistical analysis and machine learning algorithms have their limitations on analyzing metagenomics data. Thus, rest three chapters describe a new phylogeny based machine learning and feature selection algorithm to overcome these problems. The new method outperforms traditional algorithms and can provide more robust candidate microbes for further analysis. With the frowing sample size, deep neural network could potentially describe more complicated characteristic of data and thus improve model accuracy. So a deep learning framework is designed on top of the shallow learning algorithm stated above in order to further improve the prediction and selection accuracy. The present dissertation work provides a powerful tool that utilizes machine learning techniques to identify signature bacteria and key information from huge amount of metagenomics data.

This project represents an important first step in determining the fundamental aesthetically relevant features of sound. Though there has been much effort in revealing the features learned by a deep neural network (DNN) trained on visual data, little effort in applying these techniques to a network trained on audio data has been performed. Importantly, these efforts in the audio domain often impose strong biases about relevant features (e.g., musical structure). In this project, a DNN is trained to mimic the acoustic aesthetic judgment of a professional composer. A unique corpus of sounds and corresponding professional aesthetic judgments is leveraged for this purpose. By applying a variation of Google's "DeepDream" algorithm to this trained DNN, and limiting the assumptions introduced, we can begin to listen to and examine the features of sound fundamental for aesthetic judgment. / Master of Science / The question of what makes a sound aesthetically “interesting” is of great importance to many, including biologists, philosophers of aesthetics, and musicians. This project serves as an important first step in determining the fundamental aesthetically relevant features of sound. First, a computer is trained to mimic the aesthetic judgments of a professional composer; if the composer would deem a sound “interesting,” then so would the computer. During this training, the computer learns for itself what features of sound are important for this classification. Then, a variation of Google’s “DeepDream” algorithm is applied to allow these learned features to be heard. By carefully considering the manner in which the computer is trained, this algorithmic “dreaming” allows us to begin to hear aesthetically salient features of sound.

While machine learning (ML) methods have achieved considerable success on several mainstream problems in vision and language modeling, they are still challenged by their lack of interpretable decision-making that is consistent with scientific knowledge, limiting their applicability for scientific discovery applications. Recently, a new field of machine learning that infuses domain knowledge into data-driven ML approaches, termed Knowledge-Guided Machine Learning (KGML), has gained traction to address the challenges of traditional ML. Nonetheless, the inner workings of KGML models and algorithms are still not fully understood, and a better comprehension of its advantages and pitfalls over a suite of scientific applications is yet to be realized. In this thesis, I first tackle the task of understanding the role KGML plays at shaping the outputs of a neural network, including its latent space, and how such influence could be harnessed to achieve desirable properties, including robustness, generalizability beyond training data, and capturing knowledge priors that are of importance to experts. Second, I use and further develop loss landscape visualization tools to better understand ML model optimization at the network parameter level. Such an understanding has proven to be effective at evaluating and diagnosing different model architectures and loss functions in the field of KGML, with potential applications to a broad class of ML problems. / Doctor of Philosophy / My research aims to address some of the major shortcomings of machine learning, namely its opaque decision-making process and the inadequate understanding of its inner workings when applied in scientific problems. In this thesis, I address some of these shortcomings by investigating the effect of supplementing the traditionally data-centric method with human knowledge. This includes developing visualization tools that make understanding such practice and further advancing it easier. Conducting this research is critical to achieving wider adoption of machine learning in scientific fields as it builds up the community's confidence not only in the accuracy of the framework's results, but also in its ability to provide satisfactory rationale.

This thesis presents both a control scheme for naturalistic control of an exoskeleton glove and a glove design. Exoskeleton development has been focused primarily on design, improving soft actuator and cable-driven systems, with only limited focus on intelligent control. There is a need for control that is not limited to position or force reference signals and is user-driven. By implementing a motion amplification controller to increase weak movements of an impaired individual, a finger joint trajectory can be observed and used to predict their grasping intention. The motion amplification functions off of a virtual dynamical system that safely enforces the range of motion of the finger joints and ensures stability. Three grasp prediction algorithms are developed with improved levels of accuracy: regression, trajectory, and deep learning based. These algorithms were tested on published finger joint trajectories. The fusion of the amplification and prediction could be used to achieve naturalistic, user-guided control of an exoskeleton glove. The key to accomplishing this is series elastic actuators to move the finger joints, thereby allowing the wearer to deflect against the glove and inform the controller of their intention. These actuators are used to move the fingers in a nine degree of freedom exoskeleton that is capable of achieving all the grasps used most frequently in daily life. The controllers and exoskeleton presented here are the basis for improved exoskeleton glove control that can be used to assist or rehabilitate impaired individuals. / Master of Science / Millions of Americans report difficulty holding small or even lightweight objects. In many of these cases, their difficulty stems from a condition such as a stroke or arthritis, requiring either rehabilitation or assistance. For both treatments, exoskeleton gloves are a potential solution; however, widespread deployment of exoskeletons in the treatment of hand conditions requires significant advancement. Towards that end, the research community has devoted itself to improving the design of exoskeletons. Systems that use soft actuation or are driven by artificial tendons have merit in that they are comfortable to the wearer, but lack the rigidity required for monitoring the state of the hand and controlling it. Electromyography sensors are also a commonly explored technology for determining motion intention; however, only primitive conclusions can be drawn when using these sensors on the muscles that control the human hand. This thesis proposes a system that does not rely on soft actuation but rather a deflectable exoskeleton that can be used in rehabilitation or assistance. By using series elastic actuators to move the exoskeleton, the wearer of the glove can exert their influence over the machine. Additionally, more intelligent control is needed in the exoskeleton. The approach taken here is twofold. First, a motion amplification controller increases the finger movements of the wearer. Second, the amplified motion is processed using machine learning algorithms to predict what type of grasp the user is attempting. The controller would then be able to fuse the two, the amplification and prediction, to control the glove naturalistically.

The success of many computer vision problems and machine learning algorithms critically depends on the quality of the chosen distance metrics or similarity functions. Due to the fact that the real-data at hand is inherently task- and data-dependent, learning an appropriate distance metric or similarity function from data for each specific task is usually superior to the default Euclidean distance or cosine similarity. This thesis mainly focuses on developing new metric and similarity learning models for three tasks: unconstrained face verification, person re-identification and kNN classification. Unconstrained face verification is a binary matching problem, the target of which is to predict whether two images/videos are from the same person or not. Concurrently, person re-identification handles pedestrian matching and ranking across non-overlapping camera views. Both vision problems are very challenging because of the large transformation differences in images or videos caused by pose, expression, occlusion, problematic lighting and viewpoint. To address the above concerns, two novel methods are proposed. Firstly, we introduce a new dimensionality reduction method called Intra-PCA by considering the robustness to large transformation differences. We show that Intra-PCA significantly outperforms the classic dimensionality reduction methods (e.g. PCA and LDA). Secondly, we propose a novel regularization framework called Sub-SML to learn distance metrics and similarity functions for unconstrained face verifica- tion and person re-identification. The main novelty of our formulation is to incorporate both the robustness of Intra-PCA to large transformation variations and the discriminative power of metric and similarity learning, a property that most existing methods do not hold. Working with the task of kNN classification which relies a distance metric to identify the nearest neighbors, we revisit some popular existing methods for metric learning and develop a general formulation called DMLp for learning a distance metric from data. To obtain the optimal solution, a gradient-based optimization algorithm is proposed which only needs the computation of the largest eigenvector of a matrix per iteration. Although there is a large number of studies devoted to metric/similarity learning based on different objective functions, few studies address the generalization analysis of such methods. We describe a novel approch for generalization analysis of metric/similarity learning which can deal with general matrix regularization terms including the Frobenius norm, sparse L1-norm, mixed (2, 1)-norm and trace-norm. The novel models developed in this thesis are evaluated on four challenging databases: the Labeled Faces in the Wild dataset for unconstrained face verification in still images; the YouTube Faces database for video-based face verification in the wild; the Viewpoint Invariant Pedestrian Recognition database for person re-identification; the UCI datasets for kNN classification. Experimental results show that the proposed methods yield competitive or state-of-the-art performance.

One of the long-standing open problems in machine vision has been the task of ‘object segmentation’, in which an image is partitioned into two sets of pixels: those that belong to the object of interest, and those that do not. A closely related task is that of ‘parts-based object segmentation’, where additionally each of the object’s pixels are labelled as belonging to one of several predetermined parts. There is broad agreement that segmentation is coupled to the task of object recognition. Knowledge of the object’s class can lead to more accurate segmentations, and in turn accurate segmentations can be used to obtain higher recognition rates. In this thesis we focus on one side of this relationship: given the object’s class and its bounding box, how accurately can we segment it? Segmentation is challenging primarily due to the huge amount of variability one sees in images of natural scenes. A large number of factors combine in complex ways to generate the pixel intensities that make up any given image. In this work we approach the problem by developing generative probabilistic models of the objects in question. Not only does this allow us to express notions of variability and uncertainty in a principled way, but also to separate the problems of model design and inference. The thesis makes the following contributions: First, we demonstrate an explicit probabilistic model of images of objects based on a latent Gaussian model of shape. This can be learned from images in an unsupervised fashion. Through experiments on a variety of datasets we demonstrate the advantages of explicitly modelling shape variability. We then focus on the task of constructing more accurate models of shape. We present a type of layered probabilistic model that we call a Shape Boltzmann Machine (SBM) for the task of modelling foreground/background (binary) and parts-based (categorical) shapes. We demonstrate that it constitutes the state-of-the-art and characterises a ‘strong’ model of shape, in that samples from the model look realistic and that it generalises to generate samples that differ from training examples. Finally, we demonstrate how the SBM can be used in conjunction with an appearance model to form a fully generative model of images of objects. We show how parts-based object segmentations can be obtained simply by performing probabilistic inference in this joint model. We apply the model to several challenging datasets and find that its performance is comparable to the state-of-the-art.