Researchers have discovered that artificial intelligence (AI) is able to automatically discover latent units that are equivalent to real neurons in the brain’s visual processing system. This discovery has the potential to revolutionize how we understand visual perception at the single neuron level.
Understanding how our brain processes visual information is crucial to advancing AI technology. The brain’s visual cortex performs complex computations that transform millions of neural impulses from the retina into meaningful information that can be mapped to simple language.
To better understand this process, researchers collaborated with Caltech and the Chinese Academy of Science to compare single cortical neurons’ responses recorded by collaborators with “disentangling” deep neural networks. Disentangling models aim to be interpretable to humans and are trained without an external teaching signal using a self-supervised objective to reconstruct input images.
After rigorous oversight and data collection, researchers found that the disentangled units discovered by a machine-learning model called β-VAE were remarkably similar to the responses of single neurons at the end of the visual processing hierarchy in primates.
This discovery may lead to new insights on how the brain supports our ability to do visual perception and, in turn, inform the development of more efficient, transferable, and imaginative artificial intelligence systems.
In conclusion, the new findings from machine learning have the potential to feed back into the neuroscience community, shedding light on the benefit of disentangled representations for supporting intelligence in biological systems. This research could lay the groundwork for abstract reasoning or generalizable and efficient task learning.