Machine face recognition has traditionally been studied under the assumption of a carefully controlled image acquisition process. By controlling image acquisition, variation due to factors such as pose, lighting, and background can be either largely eliminated or specifically limited to a study over a discrete number of possibilities. Applications of face recognition have had mixed success when deployed in conditions where the assumption of controlled image acquisition no longer holds. This dissertation focuses on this unconstrained face recognition problem, where face images exhibit the same amount of variability that one would encounter in everyday life. We formalize unconstrained face recognition as a binary pair matching problem (verification), and present a data set for benchmarking performance on the unconstrained face verification task. We observe that it is comparatively much easier to obtain many examples of unlabeled face images than face images that have been labeled with identity or other higher level information, such as the position of the eyes and other facial features. We thus focus on improving unconstrained face verification by leveraging the information present in this source of weakly supervised data. We first show how unlabeled face images can be used to perform unsupervised face alignment, thereby reducing variability in pose and improving verification accuracy. Next, we demonstrate how deep learning can be used to perform unsupervised feature discovery, providing additional image representations that can be combined with representations from standard hand-crafted image descriptors, to further improve recognition performance. Finally, we combine unsupervised feature learning with joint face alignment, leading to an unsupervised alignment system that achieves gains in recognition performance matching that achieved by supervised alignment.

The mammalian visual system is an incredibly complex computation device, capable of performing the various tasks of seeing: navigation, pattern and object recognition, motor coordination, trajectory extrapolation, among others. Decades of research has shown that experience-dependent plasticity of cortical circuitry underlies the impressive ability to rapidly learn many of these tasks and to adjust as required. One particular thread of investigation has focused on unsupervised learning, wherein changes to the visual environment lead to corresponding changes in cortical circuits. The most prominent example of unsupervised learning is ocular dominance plasticity, caused by visual deprivation to one eye and leading to a dramatic re-wiring of cortex. Other examples tend to make more subtle changes to the visual environment through passive exposure to novel visual stimuli. Here, we use one such unsupervised paradigm, sequence learning, to study experience-dependent plasticity in the mouse visual system. Through a combination of theory and experiment, we argue that the mammalian visual system is an unsupervised learning device. Beginning with a mathematical exploration of unsupervised learning in biology, engineering, and machine learning, we seek a more precise expression of our fundamental hypothesis. We draw connections between information theory, efficient coding, and common unsupervised learning algorithms such as Hebbian plasticity and principal component analysis. Efficient coding suggests a simple rule for transmitting information in the nervous system: use more spikes to encode unexpected information, and fewer spikes to encode expected information. Therefore, expectation violations ought to produce prediction errors, or brief periods of heightened firing when an unexpected event occurs. Meanwhile, modern unsupervised learning algorithms show how such expectations can be learned. Next, we review data from decades of visual neuroscience research, highlighting the computational principles and synaptic plasticity processes that support biological learning and seeing. By tracking the flow of visual information from the retina to thalamus and primary visual cortex, we discuss how the principle of efficient coding is evident in neural activity. One common example is predictive coding in the retina, where ganglion cells with canonical center-surround receptive fields compute a prediction error, sending spikes to the central nervous system only in response to locally-unpredictable visual stimuli. This behavior can be learned through simple Hebbian plasticity mechanisms. Similar models explain much of the activity of neurons in primary visual cortex, but we also discuss ways in which the theory fails to capture the rich biological complexity. Finally, we present novel experimental results from physiological investigations of the mouse primary visual cortex. We trained mice by passively exposing them to complex spatiotemporal patterns of light: rapidly-flashed sequences of images. We find evidence that visual cortex learns these sequences in a manner consistent with efficient coding, such that unexpected stimuli tend to elicit more firing than expected ones. Overall, we observe dramatic changes in evoked neural activity across days of passive exposure. Neural responses to the first, unexpected sequence element increase with days of training while responses at other, expected time points either decrease or stay the same. Furthermore, substituting an unexpected element for an expected one or omitting an expected element both cause brief bursts of increased firing. Our results therefore provide evidence for unsupervised learning and efficient coding in the mouse visual system, especially because unexpected events drive prediction errors. Overall, our analysis suggests novel experiments, which could be performed in the near future, and provides a useful framework to understand visual perception and learning.

This work aims at developing a novel machine learning method to investigate heterogeneity in neurodevelopmental disorders, with a focus on autism spectrum conditions (ASCs). In ASCs, heterogeneity is shown at several levels of analysis, e.g., genetic, behavioral, throughout developmental trajectories, which hinders the development of effective treatments and the identification of biological pathways involved in gene-cognition-behavior links. ASC diagnosis comes from behavioral observations, which determine the cohort composition of studies in every scientific field (e.g., psychology, neuroscience, genetics). Thus, uncovering behavioral subtypes can provide stratified ASC cohorts that are more representative of the true population. Ideally, behavioral stratification can (1) help to revise and shorten the diagnostic process highlighting the characteristics that best identify heterogeneity; (2) help to develop personalized treatments based on their effectiveness for subgroups of subjects; (3) investigate how the longitudinal course of the condition might differ (e.g., divergent/convergent developmental trajectories); (4) contribute to the identification of genetic variants that may be overlooked in case-control studies; and (5) identify possible disrupted neuronal activity in the brain (e.g., excitatory/inhibitory mechanisms). The characterization of the temporal aspects of heterogeneous manifestations based on their multi-dimensional features is thus the key to identify the etiology of such disorders and establish personalized treatments. Features include trajectories described by a multi-modal combination of electronic health records (EHRs), cognitive functioning and adaptive behavior indicators. This thesis contributes in particular to a data-driven discovery of clinical and behavioral trajectories of individuals with complex disorders and ASCs. Machine learning techniques, such as deep learning and word embedding, that proved successful for e.g., natural language processing and image classification, are gaining ground in healthcare research for precision medicine. Here, we leverage these methods to investigate the feasibility of learning data-driven pathways that have been difficult to identify in the clinical practice to help disentangle the complexity of conditions whose etiology is still unknown. In Chapter 1, we present a new computational method, based on deep learning, to stratify patients with complex disorders; we demonstrate the method on multiple myeloma, Alzheimer’s disease, and Parkinson’s disease, among others. We use clinical records from a heterogeneous patient cohort (i.e., multiple disease dataset) of 1.6M temporally-ordered EHR sequences from the Mount Sinai health system’s data warehouse to learn unsupervised patient representations. These representations are then leveraged to identify subgroups within complex condition cohorts via hierarchical clustering. We investigate the enrichment of terms that code for comorbidities, medications, laboratory tests and procedures, to clinically validate our results. A data analysis protocol is developed in Chapter 2 that produces behavioral embeddings from observational measurements to represent subjects with ASCs in a latent space able to capture multiple levels of assessment (i.e., multiple tests) and the temporal pattern of behavioral-cognitive profiles. The computational framework includes clustering algorithms and state-of-the-art word and text representation methods originally developed for natural language processing. The aim is to detect subgroups within ASC cohorts towards the identification of possible subtypes based on behavioral, cognitive, and functioning aspects. The protocol is applied to ASC behavioral data of 204 children and adolescents referred to the Laboratory of Observation Diagnosis and Education (ODFLab) at the University of Trento. In Chapter 3 we develop a case study for ASCs. From the learned representations of Chapter 1, we select 1,439 individuals with ASCs and investigate whether such representations generalize well to any disorder. Specifically, we identify three subgroups within individuals with ASCs that are further clinically validated to detect clinical profiles based on different term enrichment that can inform comorbidities, therapeutic treatments, medication side effects, and screening policies. This work has been developed in partnership with ODFLab (University of Trento) and the Predictive Models for Biomedicine and Environment unit at FBK. The study reported in Chapter 1 has been conducted at the Institute for Next Generation Healthcare, Department of Genetics and Genomic Sciences, Icahn School of Medicine at Mount Sinai (NY).

Internet services, such as online shopping and chat apps, have been spreading significantly in recent years, generating substantial amounts of data. These data are precious for machine learning and consist of connections between different entities, such as users and items. These connections contain important information essential for ML models to exploit, and the need to extract this information from graphs gives rise to Graph Representation Learning. By training on these data using Graph Representation Learning methods, hidden information can be obtained, and services can be improved. Initially, the models used for Graph Representation Learning were unsupervised, such as the Deepwalk and Node2vec. These models originated from the field of Natural Language Processing. These models are easy to apply, but their performance is not satisfactory. On the other hand, while supervised models like GNN and GCN have better performance than unsupervised models, they require a huge effort to label the data and finetune the model. Nowadays, the datasets have become larger and more complex, which makes the burden heavier for applying these supervised models. A recent breakthrough in the field of Natural Language Processing may solve the problem. In the paper ‘Attention is all you need’, the authors introduce the Transformer model, which shows excellent performance in NLP. Considering that the field of NLP has many things in common with the GRL and the first supervised models all originated from NLP, it is reasonable to guess whether we can take advantage of the Transformer in improving the performance of the unsupervised model in GRL. Generating embedding for nodes in the graph is one of the significant tasks of GRL. In this thesis, the performance of the Transformer model on generating embedding is tested. Three popular datasets (Cora, Citeseer, Pubmed) are used in training, and the embedding quality is measured through node classification with a linear classification algorithm. Another part of the thesis is to finetune the model to determine the effect of model parameters on embedding accuracy. In this part, comparison experiments are conducted on the dimensions, the number of layers, the sample size, and other parameters. The experiments show that the Transformer model performs better in generating embedding than the original methods, such as the Deepwalk. Compared to supervised methods, it requires less finetuning and less training time. The characteristic of the Transformer model revealed from the experiments shows that it is a good alternative to the baseline model for embedding generation. Improvement may be made on the prepossessing and loss function of the model to get higher performance. / Internettjänster, som onlineshopping och chattappar, har spridits avsevärt de senaste åren och genererat betydande mängder data. Dessa data är värdefulla för maskininlärning och består av kopplingar mellan olika enheter, såsom användare och objekt. Dessa kopplingar innehåller viktig information som är väsentlig för ML-modeller att utnyttja, och behovet av att extrahera denna information från grafer ger upphov till Graph Representation Learning. Genom att träna på dessa data med hjälp av Graph Representation Learning-metoder kan dold information erhållas och tjänster kan förbättras. Till en början var modellerna som användes för Graph Representation Learning oövervakade, såsom Deepwalk och Node2vec. Dessa modeller härstammar från området Natural Language Processing. Dessa modeller är lätta att applicera, men deras prestanda är inte tillfredsställande. Å andra sidan, medan övervakade modeller som GNN och GCN har bättre prestanda än oövervakade modeller, kräver de en enorm ansträngning för att märka data och finjustera modellen. Numera har datamängderna blivit större och mer komplexa, vilket gör bördan tyngre för att tillämpa dessa övervakade modeller. Ett nyligen genomfört genombrott inom området Natural Language Processing kan lösa problemet. I tidningen ‘Attention is all you need’ introducerar författarna Transformer-modellen, som visar utmärkta prestanda i NLP. Med tanke på att området NLP har många saker gemensamt med GRL och att de första övervakade modellerna alla härstammar från NLP, är det rimligt att gissa om vi kan dra fördel av Transformatorn för att förbättra prestandan för den oövervakade modellen i GRL. Att generera inbäddning för noder i grafen är en av GRL:s viktiga uppgifter. I detta examensarbete testas transformatormodellens prestanda för att generera inbäddning. Tre populära datamängder (Cora, Citeseer, Pubmed) används i utbildningen, och inbäddningskvaliteten mäts genom nodklassificering med en linjär klassificeringsalgoritm. En annan del av avhandlingen är att finjustera modellen för att bestämma effekten av modellparametrar på inbäddningsnoggrannheten. I den här delen utförs jämförelseexperiment på dimensionerna, antalet lager, provstorleken och andra parametrar. Experimenten visar att Transformer-modellen presterar bättre när det gäller att generera inbäddning än de ursprungliga metoderna, såsom Deep-walk. Jämfört med övervakade metoder kräver det mindre finjustering och mindre träningstid. Den egenskap hos transformatormodellen som avslöjades från experimenten visar att den är ett bra alternativ till baslinjemodellen för inbäddningsgenerering. Förbättringar kan göras av modellens preposseing- och förlustfunktion för att få högre prestanda.

No description available.

No description available.

No description available.

<p>Graphs are the natural framework of many of today’s highest impact computing applications: from online social networking, to Web search, to product recommendations, to chemistry, to bioinformatics, to knowledge bases, to mobile ad-hoc networking. To develop successful applications in these domains, we often need representation learning methods ---models mapping nodes, edges, subgraphs or entire graphs to some meaningful vector space. Such models are studied in the machine learning subfield of graph representation learning (GRL). Previous GRL research has focused on learning node or entire graph representations through associational tasks. In this work I study higher-order (k>1-node) representations of graphs in the context of both associational and counterfactual tasks.<br> </p>

With large quantities of API logs being stored, it becomes difficult to manually inspect them and determine whether the requests are benign or anomalies, indicating incorrect access to an application or perhaps actions with malicious intent. Today, companies can rely on third-party penetration testers who occasionally attempt various techniques to find vulnerabilities in software applications. However, to be a self-sustainable company, implementing a system capable of detecting abnormal traffic which could be malicious would be beneficial. By doing so, attacks can be proactively prevented, mitigating risks faster than waiting for third parties to detect these issues. A potential solution is applying machine learning, specifically anomaly detection, which detects patterns that do not conform to normal standards. This thesis covers the process of having structured log data to find anomalies in the log data. Various unsupervised anomaly detection models were evaluated on their capabilities of detecting anomalies in API logs. These models were K-means, Gaussian Mixture Model, Isolation Forest and One-Class Support Vector Machine. The findings from the evaluation show that the Gaussian Mixture Model was the best baseline model, reaching a precision of 63%, a recall of 72%, resulting in an F1-score of 0.67, an AUC score of 0.76 and an accuracy of 0.71. By tuning the models, Isolation Forest performed the best with a precision of 67% and a recall of 80%, resulting in an F1-score of 0.73, an AUC score of 0.83 and an accuracy of 0.75. The pros and cons of each model are presented and discussed along with insights related to anomaly detection and its applicability in API log analysis and API security.

With the huge volume of biological data generated by new technologies and the booming of new machine learning based analytical tools, we expect to advance life science and human health at an unprecedented pace. Unfortunately, there is a significant gap between the complex raw biological data from real life and the data required by mathematical and statistical tools. This gap is contributed by two fundamental and universal problems in biological data that are both related to confounding effects. The first is the intrinsic complexities of the data. An observed sample could be the mixture of multiple underlying sources and we may be only interested in one or part of the sources. The second type of complexities come from the acquisition process of the data. Different samples may be gathered at different time and/or from different locations. Therefore, each sample is associated with specific distortion that must be carefully addressed. These confounding effects obscure the signals of interest in the acquired data. Specifically, this dissertation will address the two major challenges in confounding effects removal: alignment and deconvolution. Liquid chromatography–mass spectrometry (LC-MS) is a standard method for proteomics and metabolomics analysis of biological samples. Unfortunately, it suffers from various changes in the retention time (RT) of the same compound in different samples, and these must be subsequently corrected (aligned) during data processing. Classic alignment methods such as in the popular XCMS package often assume a single time-warping function for each sample. Thus, the potentially varying RT drift for compounds with different masses in a sample is neglected in these methods. Moreover, the systematic change in RT drift across run order is often not considered by alignment algorithms. Therefore, these methods cannot effectively correct all misalignments. To utilize this information, we develop an integrated reference-free profile alignment method, neighbor-wise compound-specific Graphical Time Warping (ncGTW), that can detect misaligned features and align profiles by leveraging expected RT drift structures and compound-specific warping functions. Specifically, ncGTW uses individualized warping functions for different compounds and assigns constraint edges on warping functions of neighboring samples. We applied ncGTW to two large-scale metabolomics LC-MS datasets, which identifies many misaligned features and successfully realigns them. These features would otherwise be discarded or uncorrected using existing methods. When the desired signal is buried in a mixture, deconvolution is needed to recover the pure sources. Many biological questions can be better addressed when the data is in the form of individual sources, instead of mixtures. Though there are some promising supervised deconvolution methods, when there is no a priori information, unsupervised deconvolution is still needed. Among current unsupervised methods, Convex Analysis of Mixtures (CAM) is the most theoretically solid and strongest performing one. However, there are some major limitations of this method. Most importantly, the overall time complexity can be very high, especially when analyzing a large dataset or a dataset with many sources. Also, since there are some stochastic and heuristic steps, the deconvolution result is not accurate enough. To address these problems, we redesigned the modules of CAM. In the feature clustering step, we propose a clustering method, radius-fixed clustering, which could not only control the space size of the cluster, but also find out the outliers simultaneously. Therefore, the disadvantages of K-means clustering, such as instability and the need of cluster number are avoided. Moreover, when identifying the convex hull, we replace Quickhull with linear programming, which decreases the computation time significantly. To avoid the not only heuristic but also approximated step in optimal simplex identification, we propose a greedy search strategy instead. The experimental results demonstrate the vast improvement of computation time. The accuracy of the deconvolution is also shown to be higher than the original CAM. / Doctor of Philosophy / Due to the complexity of biological data, there are two major pre-processing steps: alignment and deconvolution. The alignment step corrects the time and location related data acquisition distortion by aligning the detected signals to a reference signal. Though many alignment methods are proposed for biological data, most of them fail to consider the relationships among samples carefully. This piece of structure information can help alignment when the data is noisy and/or irregular. To utilize this information, we develop a new method, Neighbor-wise Compound-specific Graphical Time Warping (ncGTW), inspired by graph theory. This new alignment method not only utilizes the structural information but also provides a reference-free solution. We show that the performance of our new method is better than other methods in both simulations and real datasets. When the signal is from a mixture, deconvolution is needed to recover the pure sources. Many biological questions can be better addressed when the data is in the form of single sources, instead of mixtures. There is a classic unsupervised deconvolution method: Convex Analysis of Mixtures (CAM). However, there are some limitations of this method. For example, the time complexity of some steps is very high. Thus, when facing a large dataset or a dataset with many sources, the computation time would be extremely long. Also, since there are some stochastic and heuristic steps, the deconvolution result may be not accurate enough. We improved CAM and the experimental results show that the speed and accuracy of the deconvolution is significantly improved.