Human Visual Learning Inspired Effective Training Methods

It is claimed that convolutional neural networks are inspired by human vision systems. Based on the literature of development of human visual system, we know that newly born child has blurred vision initially due to rapid eye movements. This rapid eye movement is termed as Nystagmus. This paper is concerned with a novel approach to quantify the nystagmus and implementing an artificial system that can mimic the visual learning of a newly born child or person with nystagmus. To quantify the nystagmus, we have recorded 10 seconds of eye movement videos from 3 subjects and 10 trials. We estimate the eye movement frequency by tracking the eye pupil through image processing, which is then used to create a database. To simulate a suitable learning environment, we have trained our model on gradually decreasing blurriness on Dog vs Cat dataset for classification task. The novelty of the paper is in the type of training which is elicited by human visual learning system. The results show that the performance of the network improves significantly when trained iteratively with increasing level of blur.