artificial intelligence: why AI can’t replicate human vision
Computers can be taught to process incoming data, such as observing faces and cars, using artificial intelligence (AI), known as deep neural networks or deep learning. This type machine learning the process uses interconnected nodes or neurons in a layered structure resembling the human brain.
The key word is “reminds” because computers, despite the power and promise of deep learning, have not yet mastered human calculations and, most importantly, the connection and connection between the body and the brain, especially when it comes to visual recognition. study led by Marike Moore, a neuroimaging expert at Western University in Canada.
“Despite promising results, deep neural networks are far from ideal computational models of human vision,” Moore said.
Previous research has shown that deep learning cannot perfectly replicate human visual recognition, but few have tried to establish which aspects of human vision deep learning cannot emulate.
The team used a non-invasive medical test called magnetoencephalography (MEG), which measures the magnetic fields generated by the brain’s electrical currents. Using MEG data obtained from human observers while viewing the object, Moore and her team discovered one key failure point.
Discover stories that interest you
They found that easily named parts of objects like ‘eye’, ‘wheel’ and ‘face’ can explain differences in human neuronal dynamics beyond what deep learning can provide.” people can rely in part on various characteristics of an object for visual recognition and make recommendations to improve the model,” Moore said.
The study shows that deep neural networks cannot fully account for the neural responses measured in human observers when people view photos of objects, including faces and animals, and has major implications for the use of real-world deep learning models such as introspection. driving vehicles.
“This discovery provides insight into what neural networks cannot understand in images, namely visual features that indicate ecologically significant categories of objects such as faces and animals,” Moore said.
“We propose that neural networks can be improved as models of the brain by giving them a more human-like learning experience, such as a learning mode that emphasizes more strongly the behavioral pressures humans are subjected to during development.”
For example, it is important for people to quickly determine whether an object is an approaching animal or not, and if so, to predict its next sequential movement. Integrating these factors during training could benefit from the ability of deep learning approaches to model human vision.
The work was published in The Journal of Neuroscience.