A5069082023- Topic Modeling
- Face and Expression Recognition
- Natural Language Processing Techniques
- Image Retrieval and Classification Techniques
- Advanced Graph Neural Networks
- Domain Adaptation and Few-Shot Learning
- Multimodal Machine Learning Applications
- AI in cancer detection
- Text and Document Classification Technologies
- Advanced Clustering Algorithms Research
- Advanced Neural Network Applications
- Radiomics and Machine Learning in Medical Imaging
- Anomaly Detection Techniques and Applications
- Advanced Image and Video Retrieval Techniques
- Machine Learning and Data Classification
- Machine Learning in Healthcare
- Medical Image Segmentation Techniques
- Reinforcement Learning in Robotics
- Sparse and Compressive Sensing Techniques
- Medical Imaging and Analysis
- Machine Learning and Algorithms
- Text Readability and Simplification
- Video Surveillance and Tracking Methods
- Evolutionary Algorithms and Applications
- Metaheuristic Optimization Algorithms Research
Sharif University of Technology
2016-2025
Despite the vast success neural networks have achieved in different application domains, they been proven to be vulnerable adversarial perturbations (small changes input), which lead them produce wrong output. In this paper, we propose a novel method, based on gradient projection, for generating universal text; namely sequence of words that can added any input order fool classifier with high probability. We observed text classifiers are quite such perturbations: inserting even single word...
Abstract Background We developed transformer-based deep learning models based on natural language processing for early risk assessment of Alzheimer’s disease from the picture description test. Methods The lack large datasets poses most important limitation using complex that do not require feature engineering. Transformer-based pre-trained have recently made a leap in NLP research and application. These are available to understand texts appropriately, shown subsequently perform well...
Detecting various types of cells in and around the tumor matrix holds a special significance characterizing micro-environment for cancer prognostication research. Automating tasks detecting, segmenting, classifying nuclei can free up pathologists' time higher value reduce errors due to fatigue subjectivity. To encourage computer vision research community develop test algorithms these tasks, we prepared large diverse dataset nucleus boundary annotations class labels. The has over 46,000 from...
Neural sequence to text generation has been proved be a viable approach paraphrase generation. Despite promising results, paraphrases generated by these models mostly suffer from lack of quality and diversity. To address problems, we propose novel retrieval-based method for Our model first retrieves pair similar the input sentence pre-defined index. With its editor module, then editing it using extracted relations between retrieved sentences. In order have fine-grained control over process,...
Deep generative models have achieved great success in areas such as image, speech, and natural language processing the past few years. Thanks to advances graph-based deep learning, particular graph representation generation methods recently emerged with new applications ranging from discovering novel molecular structures modeling social networks. This paper conducts a comprehensive survey on learning-based approaches classifies them into five broad categories, namely, autoregressive,...
Breast cancer is the most common malignancy in women, being responsible for more than half a million deaths every year. As such, early and accurate diagnosis of paramount importance. Human expertise required to diagnose correctly classify breast define appropriate therapy, which depends on evaluation expression different biomarkers such as transmembrane protein receptor HER2. This requires several steps, including special techniques immunohistochemistry or situ hybridization assess HER2...
Sample efficiency and systematic generalization are two long-standing challenges in reinforcement learning. Previous studies have shown that involving natural language along with other observation modalities can improve sample due to its compositional open-ended nature. However, transfer these properties of the decision-making process, it is necessary establish a proper grounding mechanism. One approach this problem applying inductive biases extract fine-grained informative representations...
In recent years, there have been significant improvements in various forms of image outlier detection. However, detection performance under adversarial settings lags far behind that standard settings. This is due to the lack effective exposure scenarios during training, especially on unseen outliers, leading models failing learn robust features. To bridge this gap, we introduce RODEO, a data-centric approach generates outliers for More specifically, show incorporating (OE) and training can...
There have been several efforts to improve Novelty Detection (ND) performance. However, ND methods often suffer significant performance drops under minor distribution shifts caused by changes in the environment, known as style shifts. This challenge arises from setup, where absence of out-of-distribution (OOD) samples during training causes detector be biased toward dominant features in-distribution (ID) data. As a result, model mistakenly learns correlate with core features, using this...
Large Language Models (LLMs) struggle with hallucinations and outdated knowledge due to their reliance on static training data. Retrieval-Augmented Generation (RAG) mitigates these issues by integrating external dynamic information enhancing factual updated grounding. Recent advances in multimodal learning have led the development of Multimodal RAG, incorporating multiple modalities such as text, images, audio, video enhance generated outputs. However, cross-modal alignment reasoning...
Ensuring the reliability of Large Language Models (LLMs) in complex reasoning tasks remains a formidable challenge, particularly scenarios that demand precise mathematical calculations and knowledge-intensive open-domain generation. In this work, we introduce an uncertainty-aware framework designed to enhance accuracy LLM responses by systematically incorporating model confidence at critical decision points. We propose approach encourages multi-step LLMs quantify intermediate answers such as...
Deep learning has received much attention as of the most powerful approaches for multimodal representation in recent years. An ideal model data can reason about missing modalities using available ones, and usually provides more information when multiple are being considered. All previous deep models contain separate modality-specific networks find a shared on top those networks. Therefore, they only consider high level interactions between to joint them. In this paper, we propose framework...
Text generation is an important Natural Language Processing task with various applications. Although several metrics have already been introduced to evaluate the text methods, each of them has its own shortcomings. The most widely used such as BLEU only consider quality generated sentences and neglecting their diversity. For example, repeatedly one high sentence would result in a score. On other hand, more recent metric diversity texts known Self-BLEU ignores texts. In this paper, we propose...
In some of object recognition problems, labeled data may not be available for all categories. Zero-shot learning utilizes auxiliary information (also called signatures) describing each category in order to find a classifier that can recognize samples from categories with no instance. this paper, we propose novel semi-supervised zero-shot method works on an embedding space corresponding abstract deep visual features. We seek linear transformation signatures map them onto the features, such...
Almost all of the existing domain adaptation methods assume that test data belong to a single stationary target distribution. However, in many real world applications, arrive sequentially and distribution is continuously evolving. In this paper, we tackle problem evolving has been recently introduced. We available for source are labeled but examples can be unlabeled sequentially. Moreover, evolve over time. propose Evolving Domain Adaptation (EDA) method first finds new feature space which...
Multi-label classification has received considerable interest in recent years. classifiers usually need to address many issues including: handling large-scale datasets with instances and a large set of labels, compensating missing label assignments the training set, considering correlations between as well exploiting unlabeled data improve prediction performance. To tackle embedding-based methods represent low-dimensional space. Many state-of-the-art use linear dimensionality reduction map...
Anatomical image segmentation is one of the foundations for medical planning. Recently, convolutional neural networks (CNN) have achieved much success in segmenting volumetric (3D) images when a large number fully annotated 3D samples are available. However, rarely dataset containing sufficient segmented accessible since providing manual masks monotonous and time-consuming. Thus, to alleviate burden annotation, we attempt effectively train CNN using sparse annotation where ground truth on...