Deep learning and artificial intelligence have permeated all walks of life. In the foreseeable future, an increasing number of people will train and analyze their own deep learning models. However, the model training process involves a large amount of iterative matrix computations, and analyzing the training process is largely limited by the complexity of massive numerical calculations, and difficulty in detecting crucial training phenomena. This report introduces our technique to visualize the deep learning model training process. This technique aims to project high dimensional concepts such as representation, distances, and classification boundaries into a two-dimensional canvas, thereby facilitating model developers in their analysis and debugging efforts. Furthermore, we also propose an interactive feedback mechanism based on the visualization process to detect people's analysis intentions regarding the model training process and recommend relevant training events.

Speaker

 A/Prof. Yun Lin

 School of Electronic Information and Electrical Engineering, SJTU

Time

        2024.4.3 12:00-13:30