Do Neural Networks Show Gestalt Phenomena? An Exploration of the Law of Closure
Abstract
The Gestalt laws of perceptual organization, which describe how visual elements in an image are grouped and interpreted, have traditionally been thought of as innate despite their ecological validity. We use deep-learning methods to investigate whether natural scene statistics might be sufficient to derive the Gestalt laws. We examine the law of closure, which asserts that human visual perception tends to "close the gap" by assembling elements that can jointly be interpreted as a complete figure or object. We demonstrate that a state-of-the-art convolutional neural network, trained to classify natural images, exhibits closure on synthetic displays of edge fragments, as assessed by similarity of internal representations. This finding provides support for the hypothesis that the human perceptual system is even more elegant than the Gestaltists imagined: a single law---adaptation to the statistical structure of the environment---might suffice as fundamental.
Introduction
For nearly a century, Gestalt psychology has described "laws" of perceptual organization---rules like proximity, similarity, and closure that dictate how humans group visual elements into coherent objects. A central debate in psychology and neuroscience concerns the origin of these rules: are they innate biological primitives, or are they learned through environmental experience? While natural scene statistics have been shown to justify these laws (e.g., the statistical likelihood of collinear edges), it remained unclear if a general-purpose learner could discover these principles without explicit programming. This study investigates whether the "Law of Closure"---the tendency to perceive a whole shape from fragmented edges---emerges naturally in deep learning models trained on real-world data.
Methods
To evaluate the presence of closure in artificial systems, researchers employed a "probing" approach on the internal representations of neural networks.
Stimuli
The primary stimuli consisted of synthetic, minimalist Kanizsa-style figures. As shown in Figure 5, these included Complete triangles (d), Aligned fragments (c) which induce the illusion of a whole triangle in humans, and Disordered fragments (e) where the same components are rotated to break the illusion. The dataset comprised 992 distinct images varying in orientation, edge length, and position to ensure results were not artifacts of simple pixel-overlap.
The Closure Measure
The researchers defined a Closure Measure ($\bar{C}$) based on the cosine similarity between high-level embeddings extracted from the network's penultimate layer. A high closure score indicates that the network's internal representation of an aligned fragment is more similar to a complete triangle than to a disordered one.
Model Comparisons
The study compared a state-of-the-art InceptionV3 model against various controls:
- Untrained models: To test for architectural inductive bias.
- Degenerate data models: Networks trained on white noise or shuffled pixels.
- Shuffled label models: Networks trained to memorize random associations rather than meaningful categories.
Key Results and Insights
The experiments revealed that closure is an emergent property of training on natural images.
- Emergence from Natural Statistics: Models trained on ImageNet showed strong closure effects, whereas those trained on noise or shuffled pixels showed zero closure. This suggests that the "closure module" is not a requirement; rather, the effect emerges as a byproduct of learning to recognize objects in environments where occlusion is common.
- Architecture and Depth: While untrained ConvNets showed a weak closure effect due to local connectivity and translation invariance, training significantly amplified this. Notably, fully connected networks (FCNets) failed to show closure entirely, highlighting the necessity of spatial constraints. Furthermore, closure effects were strongest in deeper layers where receptive fields are large enough to integrate distant fragments.
- The Role of Categorization: Networks trained with shuffled labels showed weaker closure than those trained with meaningful categories. This implies that the functional task of identifying objects helps the network learn that "closing the gap" is a useful heuristic.
Discussion and Future Work
These findings suggest that the fundamental laws of perception may be an emergent byproduct of adapting to the statistical structure of our environment. However, the authors note that CNNs are not complete cognitive models. They lack the temporal dynamics and top-down feedback loops characteristic of human perception. Future research could investigate other Gestalt principles, such as "Good Continuation" or "Common Fate," and explore the impact of more biologically plausible features like recurrence and lateral inhibition.
Summary updated