AI Struggles Where Rats Excel: A New Look at Vision and Intelligence

While AI struggles with complex image transformations, rats recognize objects with remarkable ease—forcing scientists to rethink machine vision.

Research: Unraveling the complexity of rat object vision requires a full convolutional network and beyond. Image Credit: Rudmer Zwerver / Shutterstock

Rats perceive the world with a complexity that modern artificial neural networks struggle to match. This is the finding of a recent study published in the journal Patterns by the Visual Neuroscience Lab of the Scuola Internazionale Superiore di Studi Avanzati (SISSA), led by Davide Zoccolan. Using a convolutional neural network (CNN), a type of artificial intelligence particularly effective at recognizing image content, researchers attempted to replicate rats' ability to recognize objects under various conditions, altering the objects' sizes, positions, and rotations and partially obscuring them.

The results reveal that rat vision is exceptionally efficient and adaptable, even compared to advances in artificial intelligence. As the complexity of image manipulations increases, the neural network requires more resources to compete with rat discrimination ability. While mid-level layers of the CNN were sufficient for tasks involving translation, scaling, and rotation, the network needed its full depth to match rat performance when objects were partially occluded or reduced to outlines. Additionally, rats and artificial intelligence employ different image-processing strategies, suggesting that neural networks still have something to learn from neuroscience. Unlike CNNs, which rely on specific patterns for each image, rats use a consistent set of visual cues across different contexts, allowing for greater generalization.

Convolutional Neural Networks (CNN) are the most advanced tools for image recognition and are inspired, at least in part, by the functioning of the mammalian visual cortex. A CNN consists of multiple layers, each playing a specific role in the visual analysis process. The initial layers process simple image features, such as edges and contrasts, while the intermediate and final layers combine this information to recognize more complex structures and identify objects within images.

For this study, SISSA researchers carried out behavioral experiments, training rats with a reward to recognize and discriminate objects under increasingly challenging conditions. For instance, objects were rotated, resized, or partially obscured to assess both the animals' and the neural networks' ability to recognize them despite these transformations. In simpler scenarios, such as changes in position, the neural network managed to replicate the rats' accuracy using only half of the layers; however, as complexity increased, rats maintained a quite high success rate in all tests, while the network needed increasingly more layers and resources to compete, achieving comparable results only by utilizing the entire depth of the convolutional architecture. Notably, rat performance remained stable even when objects were heavily occluded or reduced to outlines—conditions where the CNN struggled until its deepest layers were engaged.

In addition, the study found considerable differences in how the neural network and the rat visual system process visual information despite the biological inspiration of the former. CNNs tended to extract features that were more screen-locked, meaning their strategies were dependent on exact positioning, whereas rats demonstrated a higher level of view invariance by relying on the same diagnostic features regardless of an object's transformation. Unlike the CNN, which relies on specific patterns for each image, rats appear to have more flexible and generalizable strategies that remain stable even when an object's appearance changes across various contexts. "Rats, often considered poor models of vision, actually display sophisticated abilities that force us to rethink the potential of their visual system and, simultaneously, the limitations of artificial neural networks," explains Davide Zoccolan. "This suggests that they could be a good model for studying human or primate visual capabilities, which have a highly developed visual cortex, even compared to artificial neural networks, which, despite their success at replicating human visual performance, often do so using very different strategies."

The study also suggests that a better understanding of the mechanisms by which rats and, more generally, mammals recognize objects through vision in complex or ambiguous settings could inspire improvements in artificial intelligence models. Simultaneously, it underscores that even the visual systems of rats, nocturnal animals that prefer other highly developed senses, such as smell, to explore the world, are quite advanced. These findings highlight the need for AI models to incorporate more biologically inspired strategies emphasizing generalizable, view-invariant object recognition rather than relying on fixed spatial patterns.