[ad_1]
By creating a computer model of a neural network that can be deceived by optical illusions such as humans, researchers have advanced knowledge about the human visual system and can help improve artificial vision.
Is this circle green or gray? Are the center lines straight or inclined?
Optical illusions can be fun to experiment and debate, but understanding how the human brain perceives these different phenomena remains an active area of scientific research. For a class of optical illusions, called contextual phenomena, we know that these perceptions depend on the context. For example, the color that you think of a central circle depends on the color of the surrounding ring. Sometimes the outer color makes the inner color look more similar, like a neighboring green ring showing a turquoise blue ring – but sometimes the outer color makes the inner color less similar, like a pink ring showing a greenish gray circle.
A team of computer vision experts from Brown University has returned to square one to understand the neural mechanisms of these contextual phenomena. Their study was published September 20 in Psychological Review.
"There is a growing consensus that optical illusions are not a bug but a feature," said Thomas Serre, associate professor of cognitive, linguistic and psychological science at Brown and senior author of the journal. "I think they're a feature. They can represent extreme cases for our visual system, but our vision is so powerful in everyday life and in recognizing objects.
For the study, the team led by Serre, which is affiliated with the Brown Carney Institute for Brain Science, began with a computational model constrained by anatomical and neurophysiological data from the visual cortex. The model aimed to capture how neighboring cortical neurons send messages to each other and adjust responses to each other when they are presented with complex stimuli such as contextual optical illusions.
One of the team's innovations in his model was a specific model of hypothetical feedback connections between neurons, Serre said. These feedback connections can increase or decrease – excite or inhibit – the response of a central neuron, depending on the visual context.
These feedback connections are not present in most deep learning algorithms. In-depth learning is a powerful type of artificial intelligence capable of learning complex patterns of data, such as image recognition and normal speech analysis. It depends on several layers of artificial neural networks. However, most deep learning algorithms involve only direct connections between layers, and not Serre's innovative feedback connections between neurons within a layer.
Once the model was built, the team presented various illusions depending on the context. The researchers "tuned" the strength of the excitatory or feedback inhibitory connections so that the model neurons respond coherently with the neurophysiological data of the visual cortex of the primates.
Then they tested the model on a variety of contextual illusions and again found that the model perceived illusions as humans.
In order to test if they made the model unnecessarily complex, they damaged the model, selectively removing some of the connections. When some connections were missing from the model, the data did not exactly match the human perception data.
"Our model is the simplest model, both necessary and sufficient to explain the behavior of the visual cortex with respect to contextual illusions," Serre said. "It was really a classic computer neuroscience work – we started with a model to explain neurophysiology data and we ended up with predictions for human psychophysical data.
In addition to providing a unifying explanation of how humans view a class of optical illusions, Serre builds on this model in an effort to improve artificial vision.
Advanced vision algorithms, such as those used to mark faces or recognize signs of stopping, have a hard time seeing the context, he noted. By including horizontal connections tuned by context-dependent optical illusions, he hopes to solve this problem.
In-depth learning programs that take the context into account may be more difficult to deceive. A sticker, when it is stuck on a stop sign, can make an artificial vision system believe that it is a speed limit sign of 65 miles per hour, which is dangerous.
Source link
