Thanks to the growth of computing resources, deep learning is used in a multitude of fields (computer vision, speech recognition, etc.) to perform all kinds of tasks (classification, prediction, text generation, etc.). However, its application is often compared to a ”black box” where it is difficult to see how a certain task is performed. This work introduces the basic concepts of deep learning and presents different architectures of deep neural net- works. It describes existing techniques implemented to visualize the parts of an image that a model considers (saliency maps) to achieve a certain task (e.g. prediction of the class of an image). This work introduces a new approach: creating a new regularization term called saliency-based regularization whose goal is to encourage the use of relevant parts of the image in order to make the prediction of a model more consistent. It also develops a method to retrieve the visualization of each neuron composing the penultimate layer of a model. These visualizations are used to compute the saliency-based regularization. The results of using this new regularization technique to train a model are encouraging and it opens the door to a new way of training a model using saliency.
|la date de réponse||26 août 2020|
|Superviseur||Benoît Frénay (Jury)|