Game-winning neural network creators head over to Google

By | January 27, 2014
This particular algorithm is fascinating not just because it outperforms all other known neural networks on the given tasks, but because the same algorithm performs well on more than just one game. That is: it isn't restricted to strong performance on just one game, and its input actions can be extended at will (they are theoretically not just limited to the buttons on the Atari console). The algorithm is fairly simple, and the learning curve stabilizes already before 100 training epochs – which is quite impressive for the algorithm.

No wonder why Google is buying DeepMind Technologies, the company behind this feat. It'll be interesting to see what Google does with them, or rather what they end up doing for Google…

/via +Russ Abbott 

Neural Net Learns Breakout Then Thrashes Human Gamers 
The end is nigh. Humans have lost another key battle in the war against computer domination

12 thoughts on “Game-winning neural network creators head over to Google

  1. Dirk Reul

    There are, of course, the obvious candidates: 

    1. Google is going to push their automated driving systems further, using advanced AI, high quality robotics and sensor tech.
    2. Google is trying to create interactive robots  using the skills from Boston Dynamics, AI creation and sensor tech. Which would make a lot of sense, especially  if combined with this little piece

    So overall Google seems to be pushing into AI and robotics creation, possibly hoping for a feedback loop of both technologies that will enable growth for both. 

    3. Terminator and Skynet 😉 

  2. Nikolai Varankine

    Humans employ 1-10 training epochs (just a speculation based on personal observations). That is likely a target. Having army of smart electronic "employees" turns business into higher profits.

  3. Dirk Reul

    This is quite fascinating. Now let's add the acquisition of Boston Dynamics and Nest and things get interesting. 


Leave a Reply

Your email address will not be published.