Rather, our results clarify which aspects of BOS can be represented by feedforward processes and which likely require additional circuitry. This progression was consistent among stimulus types and architectures, suggesting that it embodies a general attribute of hierarchical CNN architecture processing. This study shows how BOS is developed by generalization in silico hierarchical feedforward visual networks, granting insight into feedforward edge/junction detection mechanisms critical to BOS which could be analogous to biological systems.
Rows represent stimulus types, with solid indicating filled in shapes of different luminance, and contour What is a neural network indicating outlines of occluding shapes. We kept an absolute difference to ensure a strong luminance contrast that would not trivially disappear at close luminance values. The rectangle stimulus pattern provides a relatively minimal and well-controlled geometry, which leads to a reliable generation of unambiguous T-junctions and occlusion relationships. In V1, the receptive field center is defined by feedforward signals with laterally mediated near-surround and feedback-shaped far-surround areas (Angelucci and Bressloff, 2006). Residual learning is a common technique that was developed to address degradation problems in neural networks through skip connections that maintain identity mappings in early layers (He et al., 2016). Beyond simple one-to-one feedforward architectures, other modifications have been applied to achieve better performance in CNNs.
Occlusion cues based on junction geometry have been shown to support figure–ground segregation across variations in projected retinal size when relative contour support and alignment are maintained (Lesher, 1995). The configuration of T-junctions also provides a basis for ownership inference across changes in object size because their geometric patterns remain unchanged during scaling operations. This is supported by the saliency Computer Vision and Perception News map, as the deep layer activations are consistently concentrated on this fragment across stimuli and networks.
To probe whether CNNs exhibit the ability to preserve BOS assignment under fragmented conditions, we systematically designed two classes of fragmented figures to examine network performance. Nevertheless, BOS-selective neurons continue to respond when dealing with disrupted borders (Zhang and von der Heydt, 2010), which indicates the ability of biological vision to merge separate visual elements into single border representations. We tested network tolerance under fragmented border conditions, as objects in natural scenes are often displayed as discontinuous boundaries due to AI agents occlusion or fragmentation caused by shadow or texture.
PoolK, addK, mixedK are standing for networks’ layers – higher K means deeper layer. In shallow layers, networks appeared to focus on low level features such as edges, but in deeper layers, saliency was refined to T-junctions and corners (L-junctions). For each number of fragments available, all permutations of fragment positions were tested, with each point corresponding to the accuracy for a single stimulus condition with a specific fragment configuration (Figure 3C). The observed bimodal distribution arose primarily from the strong contribution of fragment 8, whose presence or absence largely determined whether sufficient occlusion-specific structure was available to support a reliable BOS decision. (B) Example of a left-owned test stimulus from the circular contour dataset, illustrating the geometric structure used in the cross-shape generalization experiment. Overall, the results demonstrated a strong directional dependence in generalization, https://aicurated.com/ with solid-trained networks transferring more effectively to contour conditions than contour-trained networks transfer to solid conditions.
This thereby determined the generalization and robustness of the networks to other types of stimuli. This design enables feature reuse and preserves low-level visual information, such as edges and corners, which are particularly critical for accurately processing our geometric stimuli and supporting reliable BOS decisions. Information flows feedforwardly through the network, with each layer operating on the output of the previous one. Labels refer to classification heads, consisting of trainable fully connected layers and sigmoid classification layers.
At MAX 2026 it promised Large language models more, with agents editing native Photoshop and Premiere files and the layers still intact, but that is due only as a private beta in early 2027. It processes visual input through stages of preprocessing, model-based detection or classification, and increasingly, contextual reasoning before producing an output. In this guide, we’ll cover what Vision AI is, how it works, where it’s being applied, what it costs to build, and where the technology is headed in 2026 and beyond.
Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher. All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. For instance, the superior accuracy and more stable saliency pattern of Inception (50 layers) during increment gap levels could arise from simply having more layers than AlexNet (8 layers). Lack of such mechanisms in purely feedforward networks like those tested most likely makes them vulnerable when contour continuity no longer supports Gestalt grouping.
]]>