Metacognitive Neural Networks

Self-regulated learning is a process of learning in which the system is guided by meta-cognitive principles ('knowing about knowing’, i.e., knowledge about when, what and how use strategies for better learning process) to planning, monitoring and evaluation of its progress against a defined goal. It has been proven that the self-regulated or meta-cognitive learning process is the best learning strategy employed by human.

Recently, we proposed many learning algorithm inspied by human cognitive learning and are called as 'Metacognitive Neural Networks’ with 'Self-regulated learning algorithm’. The learning algorithm choose appropriate strategies based on the current knowledge and evolve the architecture and adapt the parameters by answering what-to-learn, how-to-learn, and when-to-learn.

Below are some research highlights in the domain of metacognitve neural networks and self-regulated learning algorithms.

Metacognitive Neural Networks
Mathematical Optimization

In a nutshell, I seek solutions for black-box optimisation problems, be they single- or multi-objective, continuous (so far) or discrete (future interest): designing efficient algorithms with a theoretically-provable performance.

Below are some research highlights in a reverse chronological order.

  • BMOBench: A Black-box Multi-Objective Optimisation Benchmarking Platform [webpage].

  • NMSO: the Naive Multi-scale Search algorithm for expensive black-box optimisation.
    It was the second runner-up out of 28 algorithms in the BBComp’15 competition [paper, code].

  • MSO: a theoretical-analysis framework for multi-scale black-box search algorithms [paper].