AI and the Brain: How to Reveal Insights on Human Visual Processing
introduce
Understanding how the brain builds internal representations of the visual world is one of the most fascinating challenges in neuroscience. Over the past decade, deep learning has reshaped computer vision, resulting in neural networks that not only perform with human-level accuracy on recognition tasks, but also seem to process information in a way similar to our brains. This unexpected overlap raises an interesting question: Can studying AI models help us better understand how the brain itself learns to view it?
Researchers at Meta AI and école Normale Suppérieure focus on exploring this problem dinov3a self-supervised visual transformer, trained on billions of natural images. They compared the internal activation of Dinov3 with the human brain response to the same image using two complementary neuroimaging techniques. fMRI Provides high-resolution spatial maps of cortical activity, and Meg Capturing the exact timing of the brain’s response. Together, these datasets provide rich views on how the brain processes visual information.
Technical details
Research team exploration Three factors that may drive similarity to brain models: model size, the amount of training data, and the type of images used for training. To this end, the team trained multiple versions of Dinov3, changing these factors independently.
Brain model similarity
The team found strong fusion evidence during their review of Dinov3-matched brain responses. Activation of this model predicts fMRI signals in early visual areas and higher-order cortical areas. Achieving peak voxel correlation r = 0.45,Meg’s results show that the alignment time begins at the earliest 70 milliseconds after the onset of the image and lasts for three seconds. Importantly, early Dinov3 layers align with areas such as V1 and V2, while deeper layers match activities in higher order areas, including part of the prefrontal cortex.
Training Track
Tracking these similarities during training shows a development trajectory. After a small portion of training, low-level visual alignment time appears early, while advanced comparisons require billions of images. This reflects the way the human brain develops, with sensory regions maturing earlier than the associative cortex. The study shows that time alignment occurs the fastest, space alignment is slower, and coding similarities between them highlighting the hierarchical nature of representation development.
The role of model factors
The role of model factors is also explained. Larger models always achieve higher similarity scores, especially in higher-order cortical regions. Longer training improves overall alignment, and advanced representations benefit from most. The type of image is also important: human-centric image-trained models produce the strongest alignment. Those trained on satellite or cellular images showed partial convergence in early visual areas, but had weaker similarity in advanced brain areas. This suggests that ecology-related data is crucial to capturing various human-like representations.
Link to cortical properties
Interestingly, the timing of when Dinov3 represents also ranks with the structural and functional properties of the cortex. In training, areas with greater developmental expansion, thicker cortex or slower intrinsic time standards. In contrast, areas with high myelin sheaths are aligned earlier, reflecting their role in rapid information processing. These correlations suggest that AI models can provide clues about the basic biological principles of cortical tissue.
Nativism and empiricism
This study highlights the balance between innate structure and learning. Dinov3’s architecture provides it with a layered processing pipeline, but the similarity that is completely similar to the brain is only trained on ecologically valid data for a long time. This interaction between architectural priors and experience is cognitively scientifically with debates about nativism and empiricism.
Develop similarities
The similarities in human development are astonishing. Just as the sensory cortex matures rapidly in the brain and the association region develops slowly, Dinov3 is consistent with the sensory region early in training and early in the prefrontal region. This suggests that training trajectories in large-scale AI models can serve as computational analogs for the maturation of human brain function.
Beyond the visual path
The results also go beyond the traditional visual avenue. Dinov3 displays alignment in the prefrontal lobe and multimodal areas, raising questions about whether such models capture higher-order functions related to reasoning and decision-making. Although this study focused only on Dinov3, it points to the exciting possibilities of using AI as a tool to test about brain tissue and develop hypotheses.
in conclusion
In summary, this study shows that self-regulated visual models like Dinov3 are more than just powerful computer vision systems. They also approximate aspects of human visual processing, revealing how the size, training and data shapes converge between the brain and the machine. By studying how models learn to “see”, we can gain valuable insight into how the human brain itself develops its ability to perceive and interpret the world.
Check The paper is here. Check out ours anytime Tutorials, codes and notebooks for github pages. Also, please stay tuned for us twitter And don’t forget to join us 100K+ ml reddit And subscribe Our newsletter.
Michal Sutter is a data science professional with a master’s degree in data science from the University of Padua. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels in transforming complex datasets into actionable insights.
