Machine Learning aims to enable computers to learn from the world around them the same way as humans do, and it has revolutionized the world in the last few years. Deep learning, a particular method of machine learning, has played a crucial role in enabling several applications such as self-driving cars, speech-enabled user interfaces, medical image analysis, web search engines, recommendation systems, security & surveillance, object detection and recognition, and object tracking, etc. Though substantial progress has already been made in the development of machine learning algorithms and their underlying theory, several challenges remain.
For example, how to train deep models in the absence of large amounts of training data? How to build deep models for resource-constrained mobile devices? How to improve the generalization performance of deep neural networks? How to enable deep models to adapt and generalize to unknown target domains? How to protect deep models from adversarial attacks?
Our research focuses on finding answers to these problems.
The Machine Learning & Knowledge Representation (MlKr) lab focuses on theoretical and applied aspects of Machine Learning and Knowledge Representation in various domains. Machine learning deals with the study of computational methods that enable machines to learn from data.
— Data Augmentation
— Domain Adaptation
— Domain Generalization
— Hierarchical Learning
— Variational Inference
— Resource-efficient Neural Computing
— Adversarial Learning
— Quantum programming and quantum machine learning
— Knowledge Represenation
— Efficient Data Indexing
|Protasov, S., Khan, A.M., Sozykin, K. et al. Using deep features for video scene detection and annotation. SIViP 12, 991–999 (2018)||[text][demo]|
|Ahmad, M., Khan, A., Khan, A. M., Mazzara, M., Distefano, S., Sohaib, A., & Nibouche, O. (2019). Spatial prior fuzziness pool-based interactive classification of hyperspectral images. Remote Sensing, 11(9), 1136.||[text]|
|Sozykin, K., Protasov, S., Khan, A., Hussain, R., & Lee, J. (2018, June). Multi-label class-imbalanced action recognition in hockey videos via 3D convolutional neural networks. In 2018 19th IEEE/ACIS International Conference on Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing (SNPD) (pp. 146-151). IEEE.||[text]|
|Ahmad, M., Protasov, S., Khan, A. M., Hussain, R., Khattak, A. M., & Khan, W. A. (2018). Fuzziness-based active learning framework to enhance hyperspectral image classification performance for discriminative and generative classifiers. PloS one, 13(1), e0188996.||[text]|
|Khusainova, A., Khan, A., Rivera, A. R., & Romanov, V. (2021). Hierarchical Transformer for Multilingual Machine Translation. arXiv preprint arXiv:2103.03589.||[text]|
|Protasov, S. I. (2021). An approach to visual thesaurus exploration: a case study for Russian language. In ИНФОРМАТИКА: ПРОБЛЕМЫ, МЕТОДЫ, ТЕХНОЛОГИИ (pp. 1335-1342).||[text][demo]|
Aortic shunting operations need to be planned accurately. To do this, surgeons need to know exact sizes of aorta. This project creates a tool for medical imaging software, which will automate measuring aorta using computer vision and machine learning methods. Results of this project will be later used for other applications in medical image analysis.
Semantic indexing for edge devices.
Building a small, accurate index for large scale text and images datasets to be used on edge devices. Direct document recommendation, distributed indexing strategy, duplicate detection, content-based image retrieval.
Investigation the effect of domain gaps on the performance of machine learning models performances for different applications. Our first step was to measure the change in datasets and propose a Domain Adaptation technique for fixing this problem in images, which we extended to work on video datasets and text. We also proposed a new method for domain généralisation based on adversarial reconstruction loss. We are currently investigating the effect of deploying the same model for medical imaging on different MRI scans (MRI-Sequences).
Neural Machine Translation.
Our goal is to find new effective ways to leverage such additional sources as monolingual data or parallel corpora for related languages. Together with our students, we made the review of the current approaches to utilize monolingual data in low-resource machine translation.