In Nick Baker’s presentation highlighting Machilsen’s study “The Role of Vertical Mirror Symmetry in Visual Shape Detection” (2009), he explained how symmetry helps the human visual system detect and organize shapes. With the use of Gabor-element patterns with varying levels of noise, it was demonstrated that people can rapidly detect vertically symmetrical configurations even when identifying symmetry is not an explicit task. Their research findings supported the Gestalt principle of symmetry, where it acts as a pre-attentive cue for perceptual grouping. With this, the brain automatically integrates visual elements into wholes when they align symmetrically. This work highlighted how human perception depends on recognizing long-range relationships between spatially distant features, and not just on analyzing local details. Nick baker found similar findings but identified that symmetry is not the primary principle in the identification of shapes.
Thirteen years later, Sundaram et al. (2022) studied symmetry perception from a computational neuroscience perspective. In their publication “Recurrent connections facilitate symmetry perception in deep networks”, they questioned whether modern deep neural networks, models inspired by the visual cortex, could learn to perceive symmetry in the way humans do. These researchers compared several network architectures, including feedforward convolutional models (like ResNet and DenseNet), dilated convolutional models, Transformers, and recurrent convolutional LSTMs (LSTM3). Their results concluded that only the recurrent models could generalize symmetry detection to new types of images and “see” symmetry across long distances in visual space.
This comparison between human and machine vision reflects the insights of Machilsen’s study. Just as humans can detect symmetrical structure amid noise or distortion, Sundaram’s recurrent networks learned to recognize symmetry by integrating local features across time, ultimately mimicking the recurrent feedback loops that are thought to exist in biological vision. In contrast, Feedforward networks failed to develop a general notion of symmetry, overfitting to the training data and missing the global relationships that define symmetry. Ultimately, recurrence may be essential for modeling the same type of perceptual organization that Gestalt principles describe.
Both studies intersect on the idea that symmetry perception depends on a process that integrates information. Machilsen’s study demonstrated that human observers use symmetry to resolve noisy visual input into shapes, which is a rapid and automatic process. Sundaram confirmed this idea computationally, where artificial systems only achieved human-like symmetry identification when they incorporated recurrence, enabling the grouping of spatially separate image regions.
By connecting these projects, we see a continuity from Gestalt theory to learning, suggesting that the brain’s ability to perceive symmetry arises from dynamic computation rather than from static. Machilsen revealed the phenomena, while Sundaram developed the algorithmic and architectural explanation. Together, they underscore that symmetry perception is not just a visual trick, but a shortcut into how biological and artificial systems organize the world.
References:
Machilsen, Bart et al. “The role of vertical mirror symmetry in visual shape detection.” Journal of visionvol. 9, 12 11.1-11. 18 Nov. 2009, doi: 10.1167/9.12.11
Sundaram, Shobhita et al. “Recurrent connections facilitate symmetry perception in deep networks.” Scientific reports vol. 12,1 20931. 3 Dec. 2022, doi: 10.1038/s41598-022-25219-w
No comments:
Post a Comment