On 26 August, at VNUHCM–University of Science (HCMUS), doctoral researcher Phạm Minh Hoàng, from the Department of Computer Science at HCMUS, successfully defended a doctoral thesis entitled ‘Study on tensor decomposition and applications to data recovery and compression problems’. The research was conducted under the scientific guidance of Professor Hoàng Dương Tuấn and Dr Trần Thái Sơn.

In the context of increasingly diverse and multi-dimensional structured data, converting data into vectors or matrices can result in the loss of crucial structural information. The thesis focuses on investigating tensor decomposition techniques—an effective representation for multi-dimensional data—and potential applications in data recovery and parameter reduction for deep learning models.

For the data recovery problem, the thesis proposes a tensor augmentation technique that enables data representation in higher-dimensional structures, thereby improving recovery efficiency and reducing storage costs. The research also improves tensor completion algorithms to handle missing or non-uniform data more effectively, whilst simultaneously reducing computational costs.

For the problem of parameter reduction in deep learning models, the thesis applies Tucker and Tensor-Train (TT) decomposition methods to represent model parameters through smaller components. Proceeding from this approach, the research develops TT-ViT, which reduces parameter counts in Vision Transformer models, and proposes LoRAT, extending the LoRA method into the tensor domain to reduce fine-tuning parameters in convolutional layers whilst maintaining model performance.

Research findings hold potential applications across various fields, including image and video processing, recommendation systems, and the deployment of deep learning models on resource-constrained devices such as mobile devices, IoT, and embedded systems. Furthermore, the developed methods can be expanded into computer vision, natural language processing, biomedicine, and signal analysis.

With these achieved results, the thesis contributes to advancing tensor decomposition methods for data recovery, compression, and parameter reduction in deep learning models, whilst opening further research directions in handling large, complex datasets and applying these methods to practical problems. HCMUS congratulates doctoral researcher Phạm Minh Hoàng on successfully defending the doctoral thesis in Computer Science.