What Role Does Whole Shape(Configural Shape) Processing Play In Human Object Recognition Perception Systems and Why Do Artificial Neural Networks Do Not Succeed to Replicate It?

Nicholas Baker is a researcher in visual perception and computational neuroscience who has contributed a lot towards the understanding of how humans along with artificial intelligence recognize objects through shape. His work primarily focuses on the fundamental differences between human perception and deep learning models, specifically on how every system processes information on objects through shape. In the study, he expresses examination of how humans relied on the configural relationships between different parts of an object which means how the specific features are arranged relative to one another compared to only identifying isolated features local to the object(Baker et. al 2022). The concepts discussed in his research connect extremely closely to the ideas learned in class regarding perception and cognition where small changes in structure can have an impact significantly on recognition. Baker demonstrates that while humans naturally process objects as one whole, there are deep convolutional neural networks that often do not succeed to capture the configural nature, instead relying more heavily on local features or patterns. Their findings highlight that object recognition is not just primarily about identifying individual features of the object but about understanding the entire spatial relationships that bind those features into a whole. His research emphasizes that differences at the smallest levels for visual organization can have extremely significant impact on how both biological and artificial systems interpret the objects in the world.  

According to the study conducted by Nicholas Baker, human object perception relies heavily on the configural relationships between the local shape components of the object, which means that individuals recognize objects not just by isolated parts but by how those parts are spatially arranged into a coherent whole(Baker et. al 2022). In their experiments, participants were shown animal silhouettes that were either intact, broken, or rearranged referenced as the Frankenstein condition while also preserving local features on the objects. The results displayed and expressed that disrupting the overall configuration significantly causes problems with human recognition performance in which they decline even when the specific features on the object stay the same. In contrast, the deep convolutional neural networks (DCNNs), such as ResNet-50 and VGG-19 displayed little to no decrease in the performance when configural relationships were disrupted which helped indicate that the following models rely much on the local shape cues(or features) comparatively to processing as a whole(holistically). Furthermore, even when networks were retrained to emphasize shape over texture, or adjusted with more complex architectures such as recurrent connections in which they still did not succeed to exhibit human-like configural sensitivity(Baker et. al 2022). As a result, the study concludes that while humans process visual information in a whole perspective and spatially integrated manner, current deep learning models that lack the ability and instead function similarly to the systems that add up independent components in which it highlights a fundamental difference between biological and artificial vision systems that suggest future models must primarily incorporate border perceptual tasks for better replication experiments on the configural nature of human shape perception. 

Human perception of visual objects relies strongly on the ability to organize individual components into a whole which is a process that is essential for accurate recognition. A recent study by Dehn et al. investigated to recognize objects even when visual information is incomplete or altered. The researchers expressed that humans do not simply detect isolated parts of an object but primarily rely on the relationships between those parts toward a complete perception of shape. As a result, it suggests that the brain naturally prioritizes whole structure and configural relationships  compared to individual features of the object during visual processing of the objects(Dehn et. al 2025). However, in the study there is no direct examination of deep learning models, the findings are extremely consistent with those from Baker’s study in which together they highlight the fundamental differences in visual processing with humans compared to artificial intelligence. Baker displayed that humans cannot process configural relationships when they are altered. Deep neural networks remain largely unaffected which indicates a reliance on local features compared to whole processing. The similarity between these studies suggests that human perception is inherently structured around the spatial relationships between the individual features, whereas current artificial systems do not succeed to capture that level of organization(Dehn et. al 2025). However, it raises an important question: if human perception depends primarily on whole image processing, could future machine learning models be designed to process images via configural processing leading to more effective human-like visual recognition based technological systems? 

The ability to organize and interpret visual information through using separate parts then combining them to form the whole image has a significant impact on how humans understand the world such as when driving and you notice a stop sign based on the red octagonal shape without seeing the word stop.  The research from Baker demonstrates that human object recognition depends significantly on configural processing while the study from Dehn further supports that the brain naturally puts the small pieces together to form the entire image. Together both studies display that human perception is based on seeing the entire picture and not the different parts that make up the whole image. However, most artificial intelligence systems do not process the images based on a whole of parts, they focus primarily on the small, minute details. The difference expresses that understanding how humans process visual stimuli and information can help improve artificial intelligence programs in the future in accurately processing visual stimuli similarly to humans. 

References 

Baker , Nicholas, and James  H Elder. Deep Learning Models Fail to Capture the Configural Nature of Human Shape Perception, 16 Sept. 2022, www.sciencedirect.com/science/article/pii/S2589004222011853. 

Dehn, Kira  Isabel, et al. “Human Shape Perception Spontaneously Discovers the Biological Origin of Novel, but Natural, Stimuli | Journal of the Royal Society Interface | The Royal Society.” Human Shape Perception Spontaneously Discovers the Biological Origin of Novel, but Natural, Stimuli, 21 May 2025, royalsocietypublishing.org/rsif/article/22/226/20240931/235863/Human-shape-perception-spontaneously-discovers-the.