Neural network size estimation method based-on hierarchical force-directed graph drawing for multi-task learning

A neural network size estimation method for multi-task learning is proposed by visualizing neurons with their weights in network structure on tasks. It provides criteria with lower computational cost for adjusting number of neurons in each layer during a finding process of suitable network structure. It is evaluated by visualizing neural networks learned on the MNIST database of handwritten digits, and the result shows that inactive neurons, namely those that do not have close relation with any tasks, are located on the periphery part of visualized network, and that cutting half of training data on one specific task (out of ten) causes a 15% increase in the variance of neurons in clusters reacting to that specific task than it reacting to all tasks. The proposal aims to be developed to support the design process of a neural network dealing with multi-task of different categories, for example, one neural network for both vision and motion system of a robot.