Why do neural networks generalize so poorly?

Why do neural networks generalize so poorly?

  • June 13, 2018
Table of Contents

Why do neural networks generalize so poorly?

Deep convolutional network architectures are often assumed to guarantee generalization for small image translations and deformations. In this paper we show that modern CNNs (VGG16, ResNet50, and InceptionResNetV2) can drastically change their output when an image is translated in the image plane by a few pixels, and that this failure of generalization also happens with other realistic small image transformations. Furthermore, the deeper the network the more we see these failures to generalize.

We show that these failures are related to the fact that the architecture of modern CNNs ignores the classical sampling theorem so that generalization is not guaranteed. We also show that biases in the statistics of commonly used image datasets makes it unlikely that CNNs will learn to be invariant to these transformations. Taken together our results suggest that the performance of CNNs in object recognition falls far short of the generalization capabilities of humans.

Source: arxiv.org

Tags :
Share :
comments powered by Disqus

Related Posts

Americans Less Trusting of Self-Driving Safety Following High-Profile Accidents

Americans Less Trusting of Self-Driving Safety Following High-Profile Accidents

Americans are less trusting of self-driving cars following two deadly accidents involving autonomous or semi-autonomous vehicles, with half of U.S. adults considering those automobiles less safe than human drivers, according to a new poll. A Morning Consult survey conducted March 29-April 1 among a national sample of 2,202 adults found that 27 percent of respondents said self-driving cars are safer than human drivers, while 50 percent said autonomous vehicles are less safe. Eight percent said the automobiles are on par with human drivers when it comes to safety.

Read More
OpenAI: Gym Retro

OpenAI: Gym Retro

We’re releasing the full version of Gym Retro, a platform for reinforcement learning research on games. This brings our publicly-released game count from around 70 Atari games and 30 Sega games to over 1,000 games across a variety of backing emulators. We’re also releasing the tool we use to add new games to the platform.

Read More
Learn Reinforcement Learning from scratch

Learn Reinforcement Learning from scratch

Deep RL is a field that has seen vast amounts of research interest, including learning to play Atari games, beating pro players at Dota 2, and defeating Go champions. Contrary to many classical Deep Learning problems that often focus on perception (does this image contain a stop sign?) , Deep RL adds the dimension of actions that influence the environment (what is the goal, and how do I get there?).

Read More