Scientists have taught computers to see optical illusions, an advance that may help artificial vision algorithms to take context into account and be more robust.
Understanding how human brains perceive optical illusions remains an active area of research, said scientists from the Brown University in the US.
For one class of optical illusions, called contextual phenomena, those perceptions are known to depend on context.
For example, the color you think a central circle depends on is the color of the surrounding ring.
Sometimes the outer color makes the inner color appear more similar, such as a neighboring green ring making a blue ring appear turquoise.
The team started with a computational model constrained by anatomical and neurophysiological data of the visual cortex.
The model aimed to capture how neighboring cortical neurons send messages to each other and adjust one another’s responses when presented with complex stimuli such as contextual optical illusions.
These feedback connections are able to 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, researchers said.
Deep learning is a powerful kind of artificial intelligence that is able to learn complex patterns in data, such as recognizing images and parsing normal speech, they said.
It depends on multiple layers of artificial neural networks working together. However, most deep learning algorithms only include feed-forward connections between layers, not Serre’s innovative feedback connections between neurons within a layer. Once the model was constructed, the team presented it a variety of context-dependent illusions.
The researchers “tuned” the strength of the feedback excitatory or inhibitory connections so that model neurons responded in a way consistent with neurophysiology data from the primate visual cortex.
Then they tested the model on a variety of contextual illusions and again found the model perceived the illusions like humans.
In order to test if they made the model needlessly complex, they lesioned the model — selectively removing some of the connections.
When the model was missing some of the connections, the data didn’t match the human perception data as accurately.