| Deep neural networks are the current state-of-the-art in computer vision. In the first section, we apply these networks to the problem of crater detection, demonstrating that being able to learn filters end-to-end with a classifier is superior than existing techniques. Our models achieve state-of-the-art performance on a standard crater detection task. In the second section, we propose a measure of unit importance in neural networks. We demonstrate that using this measure, the unique features and locations of an image can be extracted and analyzed. Results also demonstrate some interesting properties of unit importance in neural nets, and we show several use cases of our measure on a face recognition data set. Finally, we address optimization difficulties unique to neural nets, and propose a new method of weight initialization which leads to better performance for deeper networks. |
