6:30 pm Folks arrive and enjoy pizza, salad, and adult beverages
7:00 pm Speakers do their thing
8:00 pm Hang out and socialize
Interpreting Trained Convolutional Neural Networks
Convolutional neural networks (CNN) are impressively powerful in various applications of computer vision, including image classification, face recognition, video analysis, etc. However, the understanding for why it works so well is limited. For example, even a principle for its architectures is lacking.
In this presentation, I will first go over basic concepts of neural networks and especially the convolutional neural network. Then I’ll show a trained hand-written digit classifier that achieved 99.3% accuracy on the MNIST dataset. I will show a few visualization of the intermediate excitation and discuss if the model learns translational invariance.