Deep Learning, Fast
Deep learning machines have been generating incredible amounts of buzz
in recent months. Their extensive abilities can allow them to play video
games, recognize faces, and, most importantly, learn. However, these
systems learn 10 times more slowly than humans, which has allowed us to
keep the creeping fears of a complete artificial intelligence (AI) takeover at bay. Now, Google has developed an AI that is capable of learning almost as quickly as a human being.
Claims of this advancement in speed come from Google’s DeepMind
subsidiary in London. They say that not only can their machine
assimilate and act on new experiences much more quickly than previous AI
models, it will soon reach human-level speeds.
If you’re unfamiliar with how deep learning works, it uses layers of neural networks
to locate trends or patterns in data. If one layer identifies a
pattern, that information will be sent to the next layer. This process
continues until all the information is collected.
Systems can be taught to learn differently depending on many
different variables, such as the strength of the connection between
layers. Significant changes in one layer can dramatically change how
information is transmitted in other layers, or how it is “learned.” Deep
neural networks have many layers, so when changes are present, the
process of learning can take an incredibly long time.
However, Alexander Pritzel, one of the researchers at Google
DeepMind, and his colleagues seem to have figured out a way around this
issue. They call it “neural episodic control.”
“Neural episodic control demonstrates dramatic improvements on the speed of learning for a wide range of environments,” the team told MIT Technology Review.
“Critically, our agent is able to rapidly latch onto highly successful
strategies as soon as they are experienced, instead of waiting for many
steps of optimization.” Their method mimics the processes of learning
that occur in human and animal brains, replicating what happens first in
the prefrontal cortex and then, as a backup, in the hippocampus.
Comments
Post a Comment