Convolutional Neural Networks (CNNs) have emerged as highly successful tools for image generation, recovery, and restoration. This success is often attributed to large amounts of training data.
On the contrary, a number of recent experimental results suggest that a major contributing factor to this success is that convolutional networks impose strong prior assumptions about natural images. A surprising experiment that highlights this structural bias towards simple, natural images is that one can remove various kinds of noise and corruptions from a corrupted natural image by simply fitting (via gradient descent) a randomly initialized, over-parameterized convolutional generator to this single image. While this over-parameterized model can eventually fit the corrupted image perfectly, surprisingly after a few iterations of gradient descent one obtains the uncorrupted image, without using any training data. This intriguing phenomena has enabled state-of-the-art CNN-based denoising as well as regularization in linear inverse problems such as compressive sensing.
In this talk we take a step towards explaining this experimental phenomena by attributing it to particular architectural choices of convolutional networks. We then characterize the dynamics of fitting a two layer convolutional generator to a noisy signal and prove that early-stopped gradient descent denoises/regularizes. This results relies on showing that convolutional generators fit the structured part of an image significantly faster than the corrupted portion.
Based on joint work with Paul Hand and Mahdi Soltanolkotabi.
The Information Systems Laboratory Colloquium (ISLC)
is typically held in Packard 101 every Thursday at 4:30 pm during the academic year. Coffee and refreshments are served at 4pm in the second floor kitchen of Packard Bldg.
The Colloquium is organized by graduate students Joachim Neu, Tavor Baharav and Kabir Chandrasekher. To suggest speakers, please contact any of the students.
To receive email notifications of seminars you can join the ISL mailing list.
Reinhard Heckel is a Rudolf Moessbauer assistant professor in the Department of Electrical and Computer Engineering (ECE) at the Technical University of Munich, and an adjunct assistant professor at Rice University, where he was an assistant professor in the ECE department from 2017-2019. Before that, he spent one and a half years as a postdoctoral researcher in the Department of Electrical Engineering and Computer Sciences at the University of California, Berkeley, and a year in the Cognitive Computing & Computational Sciences Department at IBM Research Zurich. He completed his PhD in electrical engineering in 2014 at ETH Zurich and was a visiting PhD student at the Statistics Department at Stanford University. Reinhard is working in the intersection of machine learning and signal/information processing with a current focus on deep networks for solving inverse problems, learning from few and noisy samples, and DNA data storage.