Inside DeepMind

 Artificial Intelligence  

Ahead of the publication in Nature of Google DeepMind's new research on advanced neural networks, the journal has released a video featuring interviews with company founder Demis Hassabis, and some of the developers behind the breakthrough.

F or those of use who are old enough, Space Invaders was once a great early video game.  Now, Google's Deep Mind is using the classic video game to train a neural network-based artificial intelligence.

In an upcoming paper titled, "Human-level control through deep reinforcement learning," the software engineers and computational neuroscientists at DeepMind explain how their systems are extending beyond Siri and image recognition.

Related articles What Does Watson Tell Us About the Philosophy of AI?  The Rise of Artificial Intelligence  Ben Goertzel On Aspects of Artificial General Intelligence

The limited domains of what AI can do now is a challenge being taken head-on by company founder Demis Hassabis and his team. The work expands on the Neural Turing Machine and is, "the work on the paper is the first example of a full system that can actually learn to master a wide range of diverse tasks," says Hassabis in the Nature video above.

"The work on the paper is the first example of a full system that can actually learn to master a wide range of diverse tasks."

Along with Space Invaders, the DeepMind system has mastered a number of video games, using only the visual information of the games as a person would. "The only way to do this is have the machines and the algorithms do it themselves, directly from the data," says Hassabis.

The video also introduces two of the key developers behind the DeepMind system, Volodymyr Mnih and Koray Kavukcuoglu.  Mnih was a student of deep learning pioneer Geoffrey Hinton (now also with Google) at the University of Toronto. Kavukcouglu was a student of Yann LeCun, working unsupervised learning of feature extractors and multi-stage architectures for object recognition. The pair explain how the DeepMind system works.

"What the system produces as an output is a prediction for how much reward it expects to get if it presses this key right now, and continues playing," explains Mnih. Mnih also explains why this makes the system better at games like Space Invaders, but less adept at maze oriented games like Pac Man.

Kavukcuoglu also explains that the present system has a limited memory function, so that it cannot remember what it has done far into the past.  This limits the planning function of the system. Deciding what to put into memory and how to use it is one aspect the DeepMind team is working on.

Is DeepMind on the path to artificial general intelligence (AGI)? 

"If we look ten years plus out, the kind of technology that we've published now and a lot further, and building up those capabilities so eventually we can have scientific advances being assisted by AI, either AI scientists or AI-assisted scientists and actually making new breakthroughs with the helo of machine learning," says Hassabis.


SOURCE  Nature Video

By 33rd Square

Embed