Artificial Intelligence Is Learning to Keep Learning

What if you stopped learning after graduation? It sounds like most machine-learning systems are trained. They have a task and they are deployed. But some computer scientists are now developing artificial intelligence that learns and adapts continuously, much like the human brain.

Machine-learning algorithms often take the form of a neural network, a large set of simple computing elements, or neurons, which communicate via connections between them and vary in strength, or "weight." Consider an algorithm designed to recognize images. If it mislabels a picture during training, the weights are adjusted. When we are reduced to a certain threshold, the weights are frozen at set values.

The new technique splits each weight into two values ​​that combine to influence how much one neuron can activate another. The first value is trained and frozen in traditional systems. But the second value continually adjusts in response to surrounding activities in the network. Critically, the algorithm also learns how to make these weights. So the neural network learns the patterns of behavior, as well as how much to modify each of these behaviors in response to new circumstances. The researchers presented their technique in July at a conference in Stockholm, Sweden.

Applying the technique, the team created a network that has learned to reconstruct half-erased photographs. In contrast, a traditional neural network would have to reconstruct the original. The researchers also have a network that is identified by the alphabet letters-which are nonuniform, unlike typed-ones-after-seeing-example.

In another task, neural networks controlled a character moving in a simple maze to find rewards. After one million trials, we had a network with the new semiadjustable weights that we could find. The static parts of the semiadjustable weights, apparently acquired the structure of the maze, as the dynamic parts learned how to adapt to new reward locations. "This is really powerful," says Nikhil Mishra, a computer scientist at the University of California, Berkeley, who was not involved in the research, "because the algorithms can adapt more quickly to new tasks and new situations, just like humans would. "

Thomas Miconi, a computer scientist at the ride-sharing company. Uber and the paper's lead author, says his team now. In this work, Miconi wants to simulate "neuromodulation," an instantaneous network adjustment of adaptability that allows humans to be infected when information is new or important.