Emergence of Straightened Representations in Robust Neural Networks

In this project, we explore how robust training techniques give rise the emergence of straightened feature representations in feedforward neural networks when processing natural movie sequences. By leveraging adversarial robustness and random smoothing, our models exhibit a notable decrease in curvature within their latent spaces—allowing for linear interpolation of temporal frames and reliable frame reconstruction.

Key Findings

• Decreased curvature (tangling of representations) in latent spaces through robust training
• Alignment with perceptual temporal straightening in human vision
• Improved predictivity of neural activity in primary visual cortex (V1)

Related Publications