I work on computer vision and efficient deep learning, with a focus on visual speech recognition.
I am completing an M.Sc. in Artificial Intelligence at Islamic Azad University, Karaj, and preparing to defend my thesis on lip reading with deep neural networks.
My research examines uncertainty in visual speech recognition (VSR) and decision making in Go. Alongside this work, I develop chatbots and web applications for travel services.
EGOSD uses an auxiliary belief module to estimate frame-level uncertainty and control online self-distillation in a lightweight VSR model, with negligible additional inference cost.
Similar lip movements can make speech difficult to distinguish. An internal belief module estimates frame-level uncertainty using predictive entropy and provides an online teaching signal. An entropy gate limits knowledge transfer when that signal is uncertain. The lightweight VSR model is trained with CTC.
The study examines when predictions from a learned belief model are reliable enough to guide a decision in Go. It considers uncertainty gating as a way to control reliance on those predictions, with an interest in model efficiency.
Research interests
Visual speech recognition
Recognizing speech from video when lip movements are ambiguous or individual frames provide little information.
Belief and uncertainty
Using uncertainty estimates to control knowledge transfer and guide decisions.
Efficient deep learning
Lightweight visual models, knowledge distillation and model optimization.
An open-source Python package that extracts lip regions from video with MediaPipe. It supports temporal smoothing, optional illumination normalization, and configurable processing on CPU or CUDA.