A pruning-based learning strategy to visualize self structureseparationin neural networks

To visualize structure separation for multiple tasks in neural networks, a learning strategy is proposed by incorporating the concept of neural pruning into the learning process. Visualization by placing the neurons in a 2D plane makes clear the task related structure in composite tasks. Experiment on a composite task using two tasks from UCI dataset show that the neural network of the composite task is composed of more than 80% neurons in each task. The proposed learning strategy aims the transfer learning of robot motion, and a partial experimental result is also referred.