Finding the Needle in the Haystack with Convolutions: on the benefits of architectural bias

Neural Information Processing Systems (NeurIPS)


Despite the phenomenal success of deep neural networks in a broad range of learning tasks, there is a lack of theory to understand the way they work. In particular, Convolutional Neural Networks (CNNs) are known to perform much better than Fully-Connected Networks (FCNs) on spatially structured data: the architectural structure of CNNs benefits from prior knowledge on the features of the data, for instance their translation invariance. The aim of this work is to understand this fact through the lens of dynamics in the loss landscape.

We introduce a method that maps a CNN to its equivalent FCN (denoted as eFCN). Such an embedding enables the comparison of CNN and FCN training dynamics directly in the FCN space. We use this method to test a new training protocol, which consists in training a CNN, embedding it to FCN space at a certain “relax time”, then resuming the training in FCN space. We observe that for all relax times, the deviation from the CNN subspace is small, and the final performance reached by the eFCN is higher than that reachable by a standard FCN of same architecture. More surprisingly, for some intermediate relax times, the eFCN outperforms the CNN it stemmed, by combining the prior information of the CNN and the expressivity of the FCN in a complementary way. The practical interest of our protocol is limited by the very large size of the highly sparse eFCN. However, it offers interesting insights into the persistence of architectural bias under stochastic gradient dynamics. It shows the existence of some rare basins in the FCN loss landscape associated with very good generalization. These can only be accessed thanks to the CNN prior, which helps navigate the landscape during the early stages of optimization.

Related Publications

All Publications

CVPR - June 18, 2021

NeuroMorph: Unsupervised Shape Interpolation and Correspondence in One Go

Marvin Eisenberger, David Novotny, Gael Kerchenbaum, Patrick Labatut, Natalia Neverova, Daniel Cremers, Andrea Vedaldi

CVPR - June 18, 2021

Discovering Relationships between Object Categories via Universal Canonical Maps

Natalia Neverova, Artsiom Sanakoyeu, Patrick Labatut, David Novotny, Andrea Vedaldi

CVPR - June 17, 2021

Connecting What to Say With Where to Look by Modeling Human Attention Traces

Zihang Meng, Licheng Yu, Ning Zhang, Tamara Berg, Babak Damavandi, Vikas Singh, Amy Bearman

DSN - June 21, 2021

Near-Realtime Server Reboot Monitoring and Root Cause Analysis in a Large-Scale System

Fred Lin, Bhargav Bolla, Eric Pinkham, Neil Kodner, Daniel Moore, Amol Desai, Sriram Sankar

To help personalize content, tailor and measure ads, and provide a safer experience, we use cookies. By clicking or navigating the site, you agree to allow our collection of information on and off Facebook through cookies. Learn more, including about available controls: Cookies Policy