Sharif Digital Repository

In recent years, deep learning models have become one of the most widely used models in the field of artificial intelligence by showing high accuracy in various applications, in some of which the accuracy and correctness of the output of the models are very important, and in case of an error, there will be a possibility of chaotic events. Along with the progress of deep learning models, attacks have also been introduced in this field that severely compromises the security of such models and affects the accuracy and correctness of their output. Data poisoning is an attack on deep learning models wherein the attacker manipulates some data examples and adds these poisoned data to the victim's... 

Domain shift is an inevitable issue when histopathological images are analyzed in a standard laboratory. This is due to the variations in tissue handling, manual procedures for sample preparation, and differences in scanners. This can result in reduced performance of machine learning algorithms trained on images from one laboratory when applied to another. In this framework, the goal of utilizing domain generalization techniques in machine learning is to develop models that perform well in different domains. This research examines various methods of dealing with the domain shift challenges within the context of detecting mitotic cells, and proposes algorithms to improve domain generalization... 

Deep learning methods are rapidly gaining traction for clinical use in digital pathology. Despite the increasing use of graph neural networks for classifying histopathology images, due to the high accuracy of these methods in this field and the introduction of new interpretability methods for these networks, current proposed solutions still face two main issues. First, there is no comprehensive framework for evaluating the effectiveness of interpretability methods for graph networks, particularly for histopathology images. Additionally, applying conventional interpretability methods to these types of networks for pathology images, with consideration of domain-specific knowledge, has been... 

In this study, we aim to learn representations from microscopic cell images that effectively capture the features of drugs affecting the cells, allowing us to identify effective drugs for treating a disease. We employ two parallel learning paths using predictive and generative models. Specifically, we have achieved a predictive model on the RxRx19a dataset that, unlike previous models, is interpretable, optimized, and robust to dras- tic changes in drug properties. Additionally, we have developed the first generative model on this dataset, which not only generates high-quality images but also discovers a meaningful latent space. This latent space divides the representation into relevant... 

Improvements in deep neural networks and their widespread use in research and practical application have raised significant concerns about the robustness of these networks against adversarial examples designed to deceive a deep network in calcu- lating the correct output with a slight change in input. Since this is an essential issue in highly sensitive applications, it is necessary to use a training method that reduces the model’s sensitivity to these changes but still preserves the accuracy. The common method for this goal is training the model with adversarial examples. In other words, adversarial examples are generated during the training, and the model is trained with them. This allows... 

Deep reinforcement learning has achieved better cumulative rewards than humans in many environments like Atari. One drawback of these methods is their data inefficiency which makes training time-consuming, and in some cases having this amount of data is infeasible. Meta reinforcement learning can use past experiences to enable agents to adapt to new tasks faster and makes neural networks to train in a short amount of time.One of the methods in meta reinforcement learning is inferring tasks which helps exploitation policy to have good performance in new tasks. There’s a need to improve exploration policy as well as exploitation policy by gaining informative transitions about the new task.... 

Detecting out-of-distribution (OOD) input samples at the inference time is a key element in the trustworthy deployment of intelligent models. While there has been a tremendous improvement in various flavors of OOD detection in recent years, the detection performance under adversarial settings lags far behind the performance in the standard setting. In order to bridge this gap, we introduce RODEO in this paper, a data-centric approach that generates effective outliers for robust OOD detection. More specifically, we first show that targeting the classification of adversarially perturbed in- and out-of-distribution samples through outlier exposure (OE) could be an effective strategy for the... 

The development of deep neural networks is making tremendous progress in various fields, including processing Image, speech processing and other areas. Despite this tremendous achivement, neural networks have a lot of computational overhead and memory access that prevent them from being used in resource-constrained devices. We also know that many neural network applications are of great importance in mobile devices, and it is desirable for us to use their power in this regard. Many efforts have been done at different levels to solve the problem of executing deep neural networks on these devices. In this research, an approach based on offloading is used in which two different small (on the... 

One of the goals of the current research is to investigate natural structures to inhibit Ras proteins that belong to the family of guanosine triphosphatase proteins. For example, one of the common mutations is Kras mutations, which are seen exclusively in pancreatic ductal adenocarcinoma. Changes in Kras protein expression are observed in 30 percent of lung cancer cases. Kras mutations occur in 35-45 percent of colon cancers, leading to drug resistance. Our work method in this research is that we collect a dataset of natural compounds for the target protein. Then, the binding energies of these structures with the receptor protein are calculated through Autodoc Vina. After that, using deep... 

Spatial transcriptomics is a new technology that, in addition to transcriptomic cell information, provides spatial information for each of the sample cells and, if possible, histological images of the cells. Despite much research on cell indexing, little research has been done on using cell spatial information to cluster cells, and existing methods can be improved. The aim of this study is to use cell spatial data to extract more information from the samples and to better identify the cell conditions in the images, leading to better clustering than current methods. In the proposed method, in order to use spatial location data and transcriptomics simultaneously, the samples are modeled using... 

The epidemic of Covid 19 has killed millions of people worldwide. Despite the efforts of scientists around the world, there is still no cure for this disease. Approval of newly designed drugs due to clinical trial periods is time-consuming and costly. For this reason, in the current emergency situation, it is important to have a solution for screening available approved drugs in order to find effective substances for this disease.High-throughput assays are a good option for such problems. In this field of research, image-based high-throughput assays are amongst the most effective and cost-effective methods that help quantify the response of treated cells by measuring cell... 

Correct diagnosis of myocardial scar has always been a major challenge due to the low resolution of cardiac magnetic resonance imaging. The use of a gadolinium-based contrast agent that reveals a scar is the solution proposed in medical science. However, there are limitations to the use of contrast agents in some patients. In recent years, studies based on deep learning techniques have been presented, trying to identify myocardial infarction with the help of images without using contrast agent. This type of diagnosis can be done with the help of different movements of healthy and damaged tissue. Due to the lack of datasets suitable for this application, in this study, real dataset were... 

Membership inference attacks are one of the most important privacy-violating attacks in machine learning, as well as infrastructure of more serious attacks such as data extraction attacks. Since membership inference attack is used as a measure to evaluate the level of privacy protection of machine learning models, different researches have investigated and provided new methods for this attack. However, the accuracy of these attacks has not been investigated on models trained with the latest techniques such as data augmentation and regularization techniques. In this research, we see that the Lira attack, the latest membership inference attack, which has much more power compared to previous... 

Due to the overall progress in the processing of imaging tissue cells, the identification and diagnosis of complex diseases using machine learning methods has become very important. Recognizing cell characteristics such as size, shape, and chromatin design is essential in determining cell type, which can be achieved through learning methods such as deep network training. Finding the nucleus or cytoplasm of cells in medical images is a small but significant part of a long process of diagnosing and treating diseases. Today, artificial intelligence has rushed to the aid of experts in this field and has increased the speed and accuracy of experts in finding these cells and their nuclei. This... 

Recent success of Deep Learning models, resulted in their evergrowing application in many fields. However these models usually require huge datasets, which can sometimes be hard to collect. One of the challenges related to medical data, is the Batch Effect; Medical data is usually gathered through multiple experiments. Each experiment might have a slightly different conditions than the other, resulting a shift in the data related to that batch. Batch effects can have more severe impact during testing time, as the shift in the data distribution could be bigger. Many methods have been proposed to reduce or remove the effect of external conditions on data distribution.Deep Learning models have... 

Histopathology images are a type of medical images that are used to diagnose a variety of diseases. One of these illnesses is the Leukemia cancer, which has four different subtypes and is diagnosed using a blood smear image. As a result of the advancement of deep learning tools, models for diagnosing various types of disease from images have been developed in recent years.In this project, one of the best models developed to diagnose four different types of disease was replicated, and it was demonstrated that, while this model achieves acceptable accuracy, its decision is not based on medically significant criteria. In the following, a general method for diagnosing the disease is proposed... 

Clustered Regularly Interspaced Short Palindromic Repeats, or in short, CRISPR is a relatively new technology that enables geneticists and medical researchers to edit parts of the genome by removing, adding, or altering parts of the DNA. Initially found in the genomes of prokaryotic organisms such as bacteria and archaea, this technology can cure many illnesses such as blindness and cancer. A significant issue for a practical application of CRISPR systems is accurately predicting the single guide RNA (sgRNA) on-target efficacy and off-target sensitivity. While some methods classify these designs, most algorithms are on separate data with different genes and cells. The lack of... 

Supervised kernel learning has been the focus of research in recent years. Although these methods are developed based on rigorous frameworks, they fail to improve the classification accuracy in real world applications. In order to find the origin of this problem, it should be noted that the kernel function represents a prior knowledge on the labeling function. Similar to other learning problem, learning this prior knowledge needs another prior knowledge. In supervised kernel learning, only naive assumptions can be used as the prior knowledge. These include minimizing the ℓ1 and ℓ2 norms of the kernel parameters.
As an alternative approach, in Semi-Supervised Learning (SSL), unlabeled... 

Autoencoder, as an essential part of many anomaly detection methods, is lacking flexibility on normal data in complex datasets. U-Net is proved to be effective for this purpose but overfits on the training data if trained by just using reconstruction error similar to other AE-based frameworks. Puzzle-solving, as a pretext task of self-supervised learning (SSL) methods, has earlier proved its ability in learning semantically meaningful features. We show that training U-Nets based on this task is an effective remedy that prevents overfitting and facilitates learning beyond pixel-level features. Shortcut solutions, however,are a big challenge in SSL tasks, including jigsaw puzzles. We propose...