Adversarial training improves linear visual representation in neural networks
Exploring perceptual straightness in learned visual representations
Machine LearningComputer Vision and Pattern Recognition
Summary
Visual information changes over time in ways that humans seem to represent more simply or 'straightened' in their brains. The authors studied how artificial neural networks represent these changes and found that certain training methods, like making networks resistant to tricky inputs (adversarial training), make these internal visual representations more linear and stable over time. However, this effect depends on the specific task the network is trained for and does not always happen in tasks like image segmentation or frame prediction. Adding brain-inspired features helps early in the process but doesn’t guarantee the whole network’s representations become straighter.
What this means in practice
- •For computer vision engineers: Improve temporal stability of classification results on changing video data by using adversarial training to encourage linear representations.
- •For autonomous vehicle developers: Enhance robustness of object recognition in moving scenes by using neural architectures with straighter visual representations that resist input changes.
Authors
Anne Harrington, Vasha DuTell, Ayush Tewari, Mark Hamilton, Simon Stent, Ruth Rosenholtz, William T. Freeman
Abstract
Humans have been shown to use a "straightened" encoding to represent the natural visual world as it evolves in time (Hénaff et al. 2019). In the context of discrete video sequences, "straightened" means that changes between frames follow a more linear path in representation space at progressively deeper levels of processing. While deep convolutional networks are often proposed as models of human visual processing, many do not straighten natural videos. In this paper, we explore the relationship between network architecture, differing types of robustness, biologically-inspired filtering mechanisms, and representational straightness in response to time-varying input; we identify strengths and limitations of straightness as a useful way of evaluating neural network representations. We find that (1) adversarial training leads to straighter representations in both CNN and transformer-based architectures but (2) this effect is task-dependent, not generalizing to tasks such as segmentation and frame-prediction, where straight representations are not favorable for predictions; and nor to other types of robustness. In addition, (3) straighter representations impart temporal stability to class predictions, even for out-of-distribution data. Finally, (4) biologically-inspired elements increase straightness in the early stages of a network, but do not guarantee increased straightness in downstream layers of CNNs. We show that straightness is an easily computed measure of representational robustness and stability, as well as a hallmark of human representations with benefits for computer vision models.