Image representation is a key component in visual recognition systems. In visual recognition problem, the solution or the model should be able to learn and infer the quality of certain visual semantics in the image. Therefore, it is important for the model to represent the input image in a way that the semantics of interest can be inferred easily and reliably. This thesis is written in the form of a compilation of publications and tries to look into the Convolutional Networks (CovnNets) representation in visual recognition problems from an empirical perspective. Convolutional Network is a special class of Neural Networks with a hierarchical structure where every layer’s output (except for the last layer) will be the input of another one. It was shown that ConvNets are powerful tools to learn a generic representation of an image. In this body of work, we first showed that this is indeed the case and ConvNet representation with a simple classifier can outperform highly-tuned pipelines based on hand-crafted features. To be precise, we first trained a ConvNet on a large dataset, then for every image in another task with a small dataset, we feedforward the image to the ConvNet and take the ConvNets activation on a certain layer as the image representation. Transferring the knowledge from the large dataset (source task) to the small dataset (target task) proved to be effective and outperformed baselines on a variety of tasks in visual recognition. We also evaluated the presence of spatial visual semantics in ConvNet representation and observed that ConvNet retains significant spatial information despite the fact that it has never been explicitly trained to preserve low-level semantics. We then tried to investigate the factors that affect the transferability of these representations. We studied various factors on a diverse set of visual recognition tasks and found a consistent correlation between the effect of those factors and the similarity of the target task to the source task. This intuition alongside the experimental results provides a guideline to improve the performance of visual recognition tasks using ConvNet features. Finally, we addressed the task of visual instance retrieval specifically as an example of how these simple intuitions can increase the performance of the target task massively. / <p>QC 20161209</p>

Recognizing visual information is crucial for many real artificial-intelligence-based applications, ranging from domestic robots to autonomous vehicles. However, the success of deep learning methods on visual recognition tasks is highly dependent on access to large-scale labeled datasets, which are expensive and cumbersome to collect. Transfer learning provides a way to alleviate the burden of annotating data, which transfers the knowledge learned from a rich-labeled source domain to a scarce-labeled target domain. However, the performance of deep learning models degrades significantly when testing on novel domains due to the presence of domain shift. To tackle the domain shift, conventional domain adaptation methods diminish the domain shift between two domains with a distribution matching loss or adversarial loss. These models align the domain-specific feature distribution and the domain-invariant feature distribution simultaneously, which is sub-optimal towards solving deep domain adaptation tasks, given that deep neural networks are known to extract features in which multiple hidden factors are highly entangled. This thesis explores how to learn effective transferable features by disentangling the deep features. The following questions are studied: (1) how to disentangle the deep features into domain-invariant and domain-specific features? (2) how would feature disentanglement help to learn transferable features under a synthetic-to-real domain adaptation scenario? (3) how would feature disentanglement facilitate transfer learning with multiple source or target domains? (4) how to leverage feature disentanglement to boost the performance in a federated system? To address these needs, this thesis proposes deep adversarial feature disentanglement: a class/domain identifier is trained on the labeled source domain and the disentangler generates features to fool the class/domain identifier. Extensive experiments and empirical analysis demonstrate the effectiveness of the feature disentanglement method on many real-world domain adaptation tasks. Specifically, the following three unsupervised domain adaptation scenarios are explored: (1) domain agnostic learning with disentangled representations, (2) unsupervised federated domain adaptation, (3) multi-source domain adaptation.

Up to 87% of the world’s wetlands have been destroyed, considerably reducing ecosystem services these wetlands once provided. More recently, many wetlands are being restored in an attempt to regain their ecosystem service. This study seeks to determine the effects of restored wetlands on local bat habitat use. Bat activity was found to be significantly higher around the wetlands when compared to distant grassy fields; however, no significant difference was found among the restored wetlands and a remote cattle farm containing multiple water features. Geospatial models of bat distribution and bat foraging were produced using machine learning that showed higher habitat suitability and foraging activity around restored wetlands than around distant grassy fields, suggesting that wetlands provide vital habitat for insectivorous bats. This study demonstrates that restored wetlands promote bat activity and bat foraging, and restoring wetlands may be a useful means of increasing natural pest control over nearby farmlands.

Statistics reveal a huge increase in cyberattacks making technology businesses more susceptible to data loss. With increasing application of machine learning in different domains, studies have been focused on building cognitive models for traffic anomaly detection in a communication network. These studies have led to generation of datasets containing network traffic data packets, usually captured using softwares like Wireshark. These datasets contain high dimensional data corresponding to benign data packets and attack data packets of known attacks. Recent research has mainly focused on developing machine learning architectures that are able to extract useful information from high dimensional datasets to detect attack data packets in a network. In addition, machine learning algorithms are currently trained to detect only documented attacks with available training data. However, with the proliferation of new cyberattacks and zero-day attacks with little to no training data available, current employed algorithms have little to no success in detecting new attacks. In this thesis, we focus on detecting rare attacks using transfer learning from a dataset containing information pertaining to known attacks.<div><br></div><div>In the literature, there is proof of concept for both classical machine learning and deep learning approaches for anomaly detection. We show that a deep learning approach outperforms explicit statistical modeling based approaches by at least 21% for the used dataset. We perform a preliminary survey of candidate deep learning architectures before testing for transferability and propose a Convolutional Neural Network architecture that is 99.65% accurate in classifying attack data packets.<br></div><div><br></div><div>To test for transferability, we train this proposed CNN architecture with a known attack and test it's performance on attacks that are unknown to the network. For this model to extract adequate information for transferability, the model requires a higher representation of attack data in the training dataset with the current attack data comprising only 20% of the dataset. To overcome the problem of small training sets, several techniques to boost the number of attack data packets are employed like a novel synthetic dataset based training and bootstrapped dataset training.<br></div><div><br></div><div>Our study results in identification of training-testing attack pairs that show high learning transferability. Most of the strong and consistent correlations are observed among Denial of Service(DoS) training-testing attack pairs. Furthermore, we propose hypotheses for model generalization. Our results are validated by a study of dataset features and attack characteristics using the Recursive Feature Elimination(RFE) algorithm. <br></div>

This empirical research study discusses how much the model’s accuracy changes when adding a new image class by using a pre-trained model with the same labels and measuring the precision of the previous classes to observe the changes. The purpose is to determine if using transfer learning is beneficial for users that do not have enough data to train a model. The pre-trained model that was used to create a new model was the Inception V3. It has the same labels as the eight different classes that were used to train the model. To test this model, classes of wild and non-wild animals were taken as samples. The algorithm used to train the model was implemented in a single class programmed in Python programming language with PyTorch and TensorBoard library. The Tensorboard library was used to collect and represent the result. Research results showed that the accuracy of the first two classes was 94.96% in training and 97.07% in validation. When training the model with a total of eight classes, the accuracy was 91.89% in training and 95.40 in validation. The precision of both classes was detected at 100% when the model solely had cat and dog classes. After adding six additional classes in the model, the precision changed to 95.82% of the cats and 97.16% of the dogs.

Notre perception est par nature multimodale, i.e. fait appel à plusieurs de nos sens. Pour résoudre certaines tâches, il est donc pertinent d’utiliser différentes modalités, telles que le son ou l’image.Cette thèse s’intéresse à cette notion dans le cadre de l’apprentissage neuronal profond. Pour cela, elle cherche à répondre à une problématique en particulier : comment fusionner les différentes modalités au sein d’un réseau de neurones ?Nous proposons tout d’abord d’étudier un problème d’application concret : la reconnaissance automatique des émotions dans des contenus audio-visuels.Cela nous conduit à différentes considérations concernant la modélisation des émotions et plus particulièrement des expressions faciales. Nous proposons ainsi une analyse des représentations de l’expression faciale apprises par un réseau de neurones profonds.De plus, cela permet d’observer que chaque problème multimodal semble nécessiter l’utilisation d’une stratégie de fusion différente.C’est pourquoi nous proposons et validons ensuite deux méthodes pour obtenir automatiquement une architecture neuronale de fusion efficace pour un problème multimodal donné, la première se basant sur un modèle central de fusion et ayant pour visée de conserver une certaine interprétation de la stratégie de fusion adoptée, tandis que la seconde adapte une méthode de recherche d'architecture neuronale au cas de la fusion, explorant un plus grand nombre de stratégies et atteignant ainsi de meilleures performances.Enfin, nous nous intéressons à une vision multimodale du transfert de connaissances. En effet, nous détaillons une méthode non traditionnelle pour effectuer un transfert de connaissances à partir de plusieurs sources, i.e. plusieurs modèles pré-entraînés. Pour cela, une représentation neuronale plus générale est obtenue à partir d’un modèle unique, qui rassemble la connaissance contenue dans les modèles pré-entraînés et conduit à des performances à l'état de l'art sur une variété de tâches d'analyse de visages. / Our perception is by nature multimodal, i.e. it appeals to many of our senses. To solve certain tasks, it is therefore relevant to use different modalities, such as sound or image.This thesis focuses on this notion in the context of deep learning. For this, it seeks to answer a particular problem: how to merge the different modalities within a deep neural network?We first propose to study a problem of concrete application: the automatic recognition of emotion in audio-visual contents.This leads us to different considerations concerning the modeling of emotions and more particularly of facial expressions. We thus propose an analysis of representations of facial expression learned by a deep neural network.In addition, we observe that each multimodal problem appears to require the use of a different merge strategy.This is why we propose and validate two methods to automatically obtain an efficient fusion neural architecture for a given multimodal problem, the first one being based on a central fusion network and aimed at preserving an easy interpretation of the adopted fusion strategy. While the second adapts a method of neural architecture search in the case of multimodal fusion, exploring a greater number of strategies and therefore achieving better performance.Finally, we are interested in a multimodal view of knowledge transfer. Indeed, we detail a non-traditional method to transfer knowledge from several sources, i.e. from several pre-trained models. For that, a more general neural representation is obtained from a single model, which brings together the knowledge contained in the pre-trained models and leads to state-of-the-art performances on a variety of facial analysis tasks.

Indiana University-Purdue University Indianapolis (IUPUI) / Medication non-adherence and non-compliance left unaddressed can compound into severe medical problems for patients. Identifying patients that are likely to become non-adherent can help reduce these problems. Despite these benefits, monitoring adherence at scale is cost-prohibitive. Social forums offer an easily accessible, affordable, and timely alternative to the traditional methods based on claims data. This study investigates the potential of medication adherence prediction based on social forum data for diabetes and fibromyalgia therapies by using transfer learning from the Medical Expenditure Panel Survey (MEPS). Predictive adherence models are developed by using both survey and social forums data and different random forest (RF) techniques. The first of these implementations uses binned inputs from k-means clustering. The second technique is based on ternary trees instead of the widely used binary decision trees. These techniques are able to handle missing data, a prevalent characteristic of social forums data. The results of this study show that transfer learning between survey models and social forum models is possible. Using MEPS survey data and the techniques listed above to derive RF models, less than 5% difference in accuracy was observed between the MEPS test dataset and the social forum test dataset. Along with these RF techniques, another RF implementation with imputed means for the missing values was developed and shown to predict adherence for social forum patients with an accuracy >70%. This thesis shows that a model trained with verified survey data can be used to complement traditional medical adherence models by predicting adherence from unverified, self-reported data in a dynamic and timely manner. Furthermore, this model provides a method for discovering objective insights from subjective social reports. Additional investigation is needed to improve the prediction accuracy of the proposed model and to assess biases that may be inherent to self-reported adherence measures in social health networks.

京都大学 / 新制・課程博士 / 博士(工学) / 甲第23615号 / 工博第4936号 / 新制||工||1771(附属図書館) / 京都大学大学院工学研究科機械理工学専攻 / (主査)教授 椹木 哲夫, 教授 松野 文俊, 教授 藤本 健治 / 学位規則第4条第1項該当 / Doctor of Philosophy (Engineering) / Kyoto University / DFAM

Retrogressive thaw slumps (RTSs) are thermokarst landforms that develop on slopes in permafrost regions when thawing permafrost causes the land surface to collapse. RTSs are an indicator of climate change and pose a threat to human infrastructure and ecosystems in the affected areas. As the availability of ready-to-use high-resolution satellite imagery increases, automatic RTS mapping is being explored with deep learning methods. We employed a pre-trained Mask-RCNN model to automatically map RTSs on Banks Island and Victoria Island in the western Canadian Arctic, where there is extensive RTS activity. We tested the model with different settings, including image band combinations, backbones, and backbone trainable layers, and performed hyper-parameter tuning and determined the optimal learning rate, momentum, and decay rate for each of the model settings. Our final model successfully mapped most of the RTSs in our test sites, with F1 scores ranging from 0.61 to 0.79. Our study demonstrates that transfer learning from a pre-trained Mask-RCNN model is an effective approach that has the potential to be applied for RTS mapping across the Canadian Arctic.

Deep learning is increasingly becoming a viable way of classifying all types of data. Modern deep learning algorithms, such as one dimensional convolutional neural networks, have demonstrated excellent performance in classifying time series data because of the ability to identify time invariant features. A primary challenge of deep learning for time series classification is the large amount of data required for training and many application domains, such as in medicine, have challenges obtaining sufficient data. Transfer learning is a deep learning method used to apply feature knowledge from one deep learning model to another; this is a powerful tool when both training datasets are similar and offers smaller datasets the power of more robust larger datasets. This makes it vital that the best source dataset is selected when performing transfer learning and presently there is no metric for this purpose. In this thesis a metric of predicting the performance of transfer learning is proposed. To develop this metric this research will focus on classification and transfer learning for human-activity-recognition time series data. For general time series data, finding temporal relations between signals is computationally intensive using non-deep learning techniques. Rather than time-series signal processing, a neural network autoencoder was used to first transform the source and target datasets into a time independent feature space. To compare and quantify the suitability of transfer learning datasets, two metrics were examined: i) average embedded signal from each dataset was used to calculate the distance between each datasets centroid, and ii) a Generative Adversarial Network (GAN) model was trained and the discriminator portion of the GAN is then used to assess the dissimilarity between source and target. This thesis measures a correlation between the distance between two dataset and their similarity, as well as the ability for a GAN to discriminate between two datasets and their similarity. The discriminator metric, however, does suffer from an upper limit of dissimilarity. These metrics were then used to predict the success of transfer learning from one dataset to another for the purpose of general time series classification. / Thesis / Master of Applied Science (MASc) / Over the past decade, advances in computational power and increases in data quantity have made deep learning a useful method of complex pattern recognition and classification in data. There is a growing desire to be able to use these complex algorithms on smaller quantities of data. To achieve this, a deep learning model is first trained on a larger dataset and then retrained on the smaller dataset; this is called transfer learning. For transfer learning to be effective, there needs to be a level of similarity between the two datasets so that properties from larger dataset can be learned and then refined using the smaller dataset. Therefore, it is of great interest to understand what level of similarity exists between the two datasets. The goal of this research is to provide a similarity metric between two time series classification datasets so that potential performance gains from transfer learning can be better understood. The measure of similarity between two time series datasets presents a unique challenge due to the nature of this data. To address this challenge an encoder approach was implemented to transform the time series data into a form where each signal example can be compared against one another. In this thesis, different similarity metrics were evaluated and correlated to the performance of a deep learning model allowing the prediction of how effective transfer learning may be when applied.