Artificial Neural Networks on High-Level Cognitive Processes

Artificial intelligence has been a field of interest for decades, however, with recent AI technological advancements, such as the introduction of new artificial neural networks, public awareness and interest in this field has increased, making AI research an increasingly popular field.  One topic receiving close attention is the capabilities of artificial neural networks (ANNs), including their efficiency, leading to their integration into customer service, healthcare, and other industries. The potential of AI models to have the same abilities as humans but with greater speed and accuracy has been a question for many; however, it has also been questioned if these artificial neural networks have the higher-level cognitive processing of humans such as abstract recognition and spatial reasoning.

In the research article, “Beyond the Contour: How 3D Cues Enhance Object Recognition in Humans and Neural Networks,” Mikayla Cutler, Luke Baumel, and their colleagues delve into the topic of artificial intelligence by analyzing the discrepancies in object recognition based on visual cues, including shape and texture, between humans and artificial neural networks. In this study, humans and networks were tested on their ability to recognize texture-substituted images with and without 3D information, as well as their recognition abilities from different viewpoints: canonical, images of objects from a typical viewpoint, and noncanonical, images of objects from atypical viewpoints. Overall, we see that the human participants identified a higher proportion of objects correctly than the artificial networks in all experiments. Another notable result from this study demonstrates that humans and ANNs correctly categorized the objects more frequently during noncanonical vs. canonical image categorization when the images contained 3D information; however, ANNs did not show an improvement when shadows, which provide benefits for inferring 3D structure, were attached to these images. This difference highlights the use of 3D cues in humans to form structural representations, which ANNs may not utilize. Researchers say, “For objects viewed from a noncanonical perspective, it may be necessary to infer the object’s volumetric 3D structure and mentally rotate the object in depth to recognize it.” These results demonstrate a gap in the capabilities of these artificial neural networks. Although able to recognize images, similarly to humans, these networks appear to lack the capability of high-level cognitive processes such as abstract recognition and spatial reasoning, leading to a lower correct categorization of objects when images contain 3D information, as well as a lower overall proportion of objects correctly identified than human participants.

In research by Bober-Irizar and Banerjee titled: “Neural networks for abstraction and reasoning,” researchers analyze the capabilities of artificial neural networks using a neural network imitating human intuition. The neural network, PeARL, Perceptual Abstraction and Reasoning Language, has been formed to solve ARC, Abstraction & Reasoning Corpus, tasks. 88 tasks were solved by DreamCoder using PeARL including transformations, such as flipping and rotating, cropping and uncropping, color manipulation, position manipulation, counting, and more. One notable task is its ability to perform morphology tasks, such as “draw border around objects image in blue (border only),” which resemble human-like abilities. This research by Bober-Irizar and Banerjee demonstrates that with the correct training, artificial neural networks can display high-level cognitive processes such as spatial reasoning and abstract recognition.

In the study done by Cutler and Baumel, as well as the study by Bober-Irizar and Banerjee, the capabilities of artificial neural networks are analyzed, specifically regarding high-level processes such as abstract and spatial processes. Research by Cutler and Baumel suggests that human perception involves much more abstract recognition processes, while Bober-Irizar and Banerjee demonstrate that artificial neural networks, while not the same as humans, can mimic human intuition and demonstrate abstract recognition processes with the proper model training. The implication of this research suggests future directions of artificial intelligence research. As this is a field of increasing interest and advancement, additional research can be done into artificial neural networks including 3D cues and high-level processes.

 

References:

Bober-Irizar, M., & Banerjee, S. (2024). Neural networks for abstraction and reasoning. Scientific Reports, 14, 27823. https://doi.org/10.1038/s41598-024-73582-7

Cutler, M., Baumel, L. D., Tocco, J., Friebel, W., Thiruvathukal, G. K., & Baker, N. (20XX). Beyond the contour: How 3D Cues Enhance Object Recognition in Humans and Neural Networks. Journal of Vision, XX(X), 1–X. http://journalofvision.org/?/?/?