Learning to extract robust handcrafted features with a single observation via evolutionary neurogenesis

Recent advances in neuroscience demonstrate that neurogenesis in the human brain results in the born of new neurons, which evolve and replace mature neurons over time. This procedure causes a gradual reduction in the number of neurons, resulting in the human brain's fast learning and thinking abilities. This paper models brain's neurogenesis procedure by combining evolutionary algorithms with the Convolutional Neural Network (CNN) framework. This paper shows the promising effect of evolutionary neurogenesis by analyzing its performance for solving the challenging problem of handcrafted feature extraction, which is the primary requirement of all intelligent machines. The proposed approach benefits from the knowledge of a pre-trained CNN that contains mature neurons to evolve a newborn convolutional neuron, via Particle Swarm Optimization (PSO), to detect corners robustly. The proposed approach requires only a single training data to train a robust interest point detection model, and can be trained in about 20 min on CPU, which is significantly faster than other learning-based approaches. Besides, the results demonstrate that the proposed corner detection module outperforms existing techniques, in terms of robustness in various conditions, for approximately 20 percent. The proposed learning strategy can be generalized to solve other problems as well