Sharif Digital Repository

Until now many domain adaptation methods have been proposed. A major limitation of almost all of these methods is their assumption that all test data belong to a single stationary target distribution and a large amount of unlabeled data is available for modeling this target distribution. In fact, in many real world applications, such as classifying scene image with gradually changing lighting and spam email identification, data arrives sequentially and the data distribution is continuously evolving. In this thesis, we tackle the problem of adaptation to a continuously evolving target domain that has been recently introduced and propose the Evolving Domain Adaptation (EDA) method to classify... 

Due to the recent advances in the learning of multimodal data, humans tend to use computer systems in order to solve more complex problems. One of them is Visual Question Answering (VQA), where the goal is finding the answer of a question asked about the visual contents of a given image. This is an interdisciplinary problem between the areas of Computer Vision, Natural Language Processing and Reasoning. Because of recent achievements of Deep Neural Networks in these areas, recent works used them to address the VQA task. In this thesis, three different methods have been proposed which adding each of them to existing solutions to the VQA problem can improve their results. First method tries to... 

In recent years, neural networks have been widely used in most machine learning domains. However, it has been shown that these networks are vulnerable to adversarial examples. adversarial examples are small and imperceptible perturbations applied to the input which lead to producing wrong output and thus, fooling the network. This will become an important issue in security related applications of deep neural networks, such as self-driving cars and medical diagnostics. Since, in the wort-case scenario, even human lives could be threatened. Although, many works have focused on crafting adversarial examples for image data, only a few studies have been done on textual data due to the existing... 

Liquid biopsy includes information about the progress of the tumor, the effectiveness of the treatment and the possibility of tumor metastasis. This type of biopsy obtains this information by doing diagnosis and enumerating genetic variations in cells and cell-free DNA (cfDNA). Only a small fraction of cfDNA which might be free circulation tumor DNA (ctDNA) fragments, has mutations and is usually identified by epigenetic variations. On the other hand, the use of liquid biopsy has decreased, and tumors in the final stages are often untreatable due to the low accuracy in prediction of cancer. In this research, the aim is to predict cancer using cfDNA methylation patterns. We obtain these... 

Image segmentation in medical imaging, as one of the most important branches of medical image analysis, often faces the challenge of limited labeled data for application in deep learning methods. The high cost of data collection and the need for expertise in image segmentation, particularly in three-dimensional images such as MRI and CT or sequence images like CMR, have all contributed to this problem, even for popular networks like U-Net, which struggle to achieve high accuracy. As a result, research efforts have focused on semi-supervised learning approaches, weakly supervised learning, as well as multi-instance learning in medical image segmentation. Unfortunately, each of these methods... 

A predefined assumption in many learning algorithms is that the training and test data must be in thesame feature space and have the same distribution.However, this assumption may not hold in all of these algorithms and in the real world there might be difference between the source and the targer domian, whether in the feature space or the distribution. Moreover, there might be a few number of labled data of the target domain which causes difficulty in learning an accurate classifier. In such cases, transferring knowledge can be useful if can be done successfully and transfer learning was introduced for this purpose. Domain Adaptation is one of the transfer leaning problems that assume some... 

Despite recent advancements in deep learning methods, these methods rely on a huge amount of training data to work. Recently the problem of solving classification and recently semantic segmentation problems with a few training data have gained attention to tackle this issue. In this research, we propose a meta-learning method by combining optimization-based and prototypical approaches in which a small portion of parameters are optimized with task-specific initialization. In addition to this and designing other parts of the method, we propose a new approach to use query data as an unlabeled sample to enhance task-specific learning. Alongside the mentioned method, we propose an approach to use... 

With the popularity of smart devices, a lot of applications have developed and deployed.Developers try to establish continuous interaction with their users by different tools including push notifications. Push notification is a message that is sent from developers to the users and as soon as the user’s device receives that, it appears on the device screen. Sending proper content to users in order to resume their engagement is one of the most important usages of notifications. Users are not interested in receiving irrelevant notifications, and receiving irrelevant notifications make them remove the application, so it’s important to predict users’ interest in different notifications and push... 

Graphs are widely used for representing structured data and analysis of them is an important area that appears in a broad domain of applications. Graph processing is of great importance in analyzing and predicting social media users' behavior, examining financial markets, detecting malware programs, and designing recombinant drugs. For example, consider a graph in which nodes and edges show the financial institutions and the financial connection between these institutions, respectively. Financial connection refers to the investment of one institute by another. Based on the graph structure, predicting trade stability and balance is extremely significant in macro decisions.In the last few... 

Sample efficiency and systematic generalization are two long-standing challenges in sequential decision-making problems, especially, in reinforcement learning settings. It is hypothesized that involving natural language in conjunction with other observation modalities in decision-making environments can improve generalization due to its compositional and open-ended nature, and sample efficiency due to the concise information summarized in relatively short linguistic units. By exploiting this information and the compositional structure of the language, one can achieve an abstract and factored understanding of the environment and the task at hand. To do so, it is necessary to find the proper... 

One of the main challenges in the field of Machine Learning, which has come to attention in recent years, is the models' inability to generalize well to datapoints that come from a distribution different from the one that the model has been trained on. This problem is known by Out-of-Distribution Generalization. This problem is of utmost importance as in many real-world applications the training and test data do not come from the same distribution, and thus the model will fail on the test data despite its acceptable performance on samples from the same distribution of the training set. One of the main reasons behind such failure is models' reliance on spurious correlations between some... 

Few-shot learning methods have achieved notable performance in recent years. However, fewshot learning in large-scale settings with hundreds of classes is still challenging. In this dissertation, we tackle the problems of large-scale few-shot learning by taking advantage of pre-trained foundation models. We recast the original problem in two levels with different granularity. At the coarse-grained level, we introduce a novel object recognition approach with robustness to sub-population shifts. At the fine-grained level, generative experts are designed for few-shot learning, specialized for different superclasses. A Bayesian schema is considered to combine coarse-grained information with... 

Regarding the ever-increasing usage of text generation and analysis in Natural Language Processing field, Language Modeling and Masked Language Modeling have been recently one of the most frequent tasks. Besides, many pretrained models such as BERT have been proposed due to the lack of rich datasets and computational resources among researchers. These models can be finetuned on other datasets in downstream tasks. Although these Transformer-based deep neural networks have performed perfectly in many problems, they still have some shortcomings in a few tasks.Furthermore, structured data like graphs have been recently used extensively in Natural Language Processing and researchers have taken... 

Despite the remarkable success of deep learning algorithms in recent years, it still has a long way to reach the status of human natural intelligence and to acquire the expected self-autonomy. As a result, many researchers in this field have focused on the development of these algorithms while taking inspiration from human cognitive behaviors. One of the disadvantages of current algorithms is the lack of their ability to learn in a continual manner while deployed in the environment. More precisely, deep learning models are not able to gradually gather knowledge from the environment and if they are in a situation of limited access to data, they will suffer from catastrophic forgetting; a... 

Recent advances in deep neural networks have shown significant potential; however, they still face challenges when it comes to non-stationary environments. Continual learning is related to deep neural networks with limited capacity that should perform well on a sequence of tasks. On the other hand, studies have shown that neural networks are sensitive to covariate shifts. But in many cases, the distribution of data varies with time. Domain Adaptation tries to improve the performance of a model on an unlabeled target domain by using the knowledge of other related labeled data coming from a different distribution. Many studies on domain adaptation have optimistic assumptions that are not... 

Increasing the popularity of smartphones and the great ovation of users of mobile apps has turned the app stores to massive software repositories. Therefore, using these repositories can be useful for improving the quality of the program. Since the bridge between users and developers of mobile apps is the comments that users write in the app store, special attention to these comments from developers can make a dramatic improvement in final quality of mobile apps. Hence, in recent years, numerous studies have been conducted around the topic of opinion mining, whose intention was to extract and exert important information from user’s reviews. One of the shortcomings of these studies is the... 

Representation learning refers to mapping the input data to another space, usually with lower dimensions than the input space. This task can be helpful in improving the performance of methods in downstream tasks, compression, and improving sample generation in generative models. Representation learning is a problem connected to information theory, and information theory's concepts and quantities are used widely in representation learning models. Besides, the representation learning problem is closely related to latent variable generative models. These models usually learn useful representations in their process of training, implicitly or explicitly. So, the usage of latent variable... 

The existing continual learning methods are mainly focused on fully-supervised scenarios and are still not able to take advantage of unlabeled data available in the environment. Some recent works tried to investigate semi-supervised continual learning (SSCL) settings in which the unlabeled data are available, but it is only from the same distribution as the labeled data. This assumption is still not general enough for real-world applications and restricts the utilization of unsupervised data. In this work, we introduce Open-Set Semi-Supervised Continual Learning (OSSCL), a more realistic semi-supervised continual learning setting in which out-of-distribution (OoD) unlabeled samples in the... 

Forecasting and analysing financial markets has always been an interesting research topic for fields ranging from financial sciences to mathematics and statistics. With the rapid development of artificial intelligence in the recent years, there has been a growing interest in using deep neural networks to predict market future trends. The price in these markets is determined by mechanisms of demand and supply. When there is a tendancy to buy a stock, there will be an increase in demand resulting a positive growth for price. On the other hand, when a large group of investors decide to sell their assets, market will experience an increase in supply and subsequently the prices drop. Availability... 

Reinforcement learning is a field of machine learning which is more similar to human training procedures.It uses reward signals to train an agent designed to act in that environment. Deep neural networks enhance the agent’s ability to determine and act better in its complex environment. Most previous works have addressed model-free agents, which ignore modeling details of the environment that in turn can be used to achieve better results. On the other hand, humans utilize a model-based approach in their decision-making process. They use their knowledge to predict the future and choose the action that leads them to a better state. To combine the benefits of model-based and model-free designs,...