It seems we really are getting closer to artificial intelligence (AI) - and Google is leading the way. The tech giant has just revealed that its DeepMind AI can learn independently.
It appears the AI can now "learn" new things based on what's stored in its memory. Programmers from Alphabet, Google's parent company and developers of DeepMind, announced it can "intelligently build" on its memory.
This breakthrough research is now known as a Differential Neural Computer (DNC). This hybrid system pairs a a neural network with conventional computers, and DeepMind is "intelligent" enough to navigate around them.
DNC Architecture (from DeepMind)
In turn, DeepMind can scour information from these "data banks" and infer new things on its own. Researchers from Alphabet said the DNC will allow DeepMind to function more like a human brain
than its other AI counterparts.
(And did you know DeepMind has already beat a lot of Go champions? Go is considered one of the hardest mental games to master.)
DeepMind: Breaking through barriers
The DNC is demonstrating its capability to combine external memory with an AI's neural network approach. This is similar to how a mind (which a neural network tries to simulate) interacts with an external hard drive.
DeepMind researchers Alexander Graves and Greg Wayne explained DNC allows DeepMind not only to store massive amounts of data, but to learn from them as well! See a sample below.
According to the DeepMind blog post
, the DNC makes use of a "controller" that "learns" over time. It compares the results of its investigations with the correct responses it needs. Over time, this controller will be able to gather more accurate results by optimizing its memory banks.
In a demonstration, the DNC can figure out a family tree by just giving it few bits of information. It can "write, rewrite" and optimize its memory to correlate the proper information at the same time.
Or, for instance, giving it basic information from a public transit system like the London Underground allows it to understand its more complex parts without any help at all! It only has to use what it has on its memory banks.
This is akin to how smartphone maps analyze data from its internal maps. However, DeepMind will figure this out on its own by juggling millions of bits of data all at once. This is highly unlike smartphone maps that rely on its internal memory.
This breakthrough with the DNC means DeepMind is on its way to a future where AI can finally think like humans. It may be able to create and generate responses from prior knowledge, instead of dumping everything at once.