Rats see the world in ways that artificial neural networks find hard to copy. Researchers at the Visual Neuroscience Lab of SISSA reported these findings in a paper published in Patterns. They used a convolutional neural network (CNN) to see how well it could mimic rat vision. They changed the size, position, and rotation of objects or hid parts of them to test both rats and the neural network.

A CNN is a tool for recognizing images, inspired by how mammals see. It has layers that do different jobs. Early layers look at simple things like lines or color changes, while later layers put these together to recognize whole objects.

SISSA researchers trained rats with rewards to pick out objects under tough conditions. When objects were just moved around, the neural network could do as well as rats with fewer layers. But as tasks got harder, rats did much better, staying accurate even when the network needed all its layers to compete.

AI has still much to learn from brain science

The study also showed that rats process images differently from how CNNs do. Rats seem to use more adaptable strategies, keeping their recognition skills strong even when objects look different. This suggests that rats, often seen as having poor vision, actually have complex visual abilities that could help us understand human or primate vision better.

In a SISSA press release, research leader Davide Zoccolan notes that rats might be good models for studying visual recognition in more advanced species like humans, whose visual cortex is very developed. Even though artificial intelligence (AI) can mimic human vision, it often does so differently than biological systems do. Learning from how rats see could help improve AI. Rats, despite being nocturnal and relying more on smell, have surprisingly advanced vision.

The study shows that rat vision is very effective and flexible. As the images got more complicated, the neural network needed more resources to keep up with the rats. Rats and AI use different methods to understand images, which means there’s still much AI can learn from brain science.

Let us know your thoughts! Sign up for a Mindplex account now, join our Telegram, or follow us on Twitter.