Computational Intelligence Research Group (CIRG)

Bangla is an important language with a rich heritage; 21st February is declared as the International Mother Language day by UNESCO. Although it is the fifth most spoken language globally (spoken by ~245 million people), systematic efforts for its computerization are critical. CIRG is actively working on tasks including Language Corpus development, Optical Character Recognition (OCR), Text-to-Speech (TTS), and Speech-to-Text (STT) systems.

In machine learning, ensemble methods use multiple models to obtain superior predictive performance compared to constituent individual models. Ensembles promote diversity among models (such as neural networks and decision trees) to yield robust results. We focus on developing advanced ensemble approaches for enhanced performance.

Modeled after the human brain, artificial neural networks are efficient tools in computational intelligence. We research automatic architecture learning for better generalization and extend models to handle complex-valued data in signal processing and imaging. Additionally, we study fuzzy systems for imprecise information processing and hybrid neuro-fuzzy control techniques.

Natural systems inspire computer algorithms designed to tackle complex optimization. Swarm Intelligence (SI) simulates collective behavior (e.g., ant colonies, bird flocks) without centralized control to solve distributed problem-solving tasks and real-world optimization challenges.

Evolutionary Computation (EC) uses population-based metaheuristic optimization algorithms inspired by biological evolution (reproduction, mutation, recombination). We develop novel EC optimization algorithms and apply them to solve complex real-life optimization problems.

Machine pattern recognition is a core goal in artificial intelligence. Our research targets recognition challenges in Bangla speech and text processing, crafting hybrid frameworks, and exploring foundational theoretical aspects.

Feature representation is essential for applying machine learning algorithms. While traditional methods rely on hand-engineering features, unsupervised feature learning automates this process to extract optimal representations directly from unlabeled data, helping overcome current limitations in modern machine learning paradigms.