Varga, Dániel (Rényi Institute of
Mathematics)
Neural Art
In this talk we provide a nontechnical, visual introduction to recent generative deep learning techniques for vision tasks. These algorithms can create or alter images in powerful and often surprising ways. The first part of the talk is dedicated to optimization algorithms styled after Google's popular Deep Dream algorithm. These algorithms start from a trained neural network and an input image, and iteratively modify the image until they achieve some well-chosen neural firing pattern. As a simplest example, the following three images were optimized to strongly excite three specific neurons of an image recognition network. This technique is useful for peeking into artificial neural networks, models that were traditionally considered notoriously opaque. We'll show how this idea can be extended to solve the style transfer task (as recently popularized by the Prisma mobile app), and even so-called image analogy tasks: In the second part of the talk we introduce another class of deep learning models called generative autoencoders. Their mathematics is especially attractive, and they are competent at visual analogy tasks like modifying face images to alter facial features, age, or facial expression:
(Images by Alexander Mordvintsev.)
(The fourth image is calculated from the first three images. Image by Adam Wentz.)
(Image by Tom White.)
Date: Dec. 6, Tuesday 4:15pm
Place: BME, Building „Q”, Room QBF13