[Tyler]

How Google Aims To Dominate Artificial Intelligence
9 November 2015, 12:00 am

                    Okay, so I mentioned that Google Now works with KnowledgeGraph. What's that?

Popular ScienceLink

Source: AI in the News

To visit any links mentioned please view the original article, the link is at the top of this post.

[ivan.moony]

So they know how to recognize pictures and speech, but they are stuck with NLP for better Google search. I wonder why don't they even mention deduction and induction? I think those two are crucial for understanding texts.

[Don Patrick]

Google's main rowboat is neural nets, who are great at finding average patterns in repetitive data. e.g. show a neural network 1 million images of cats and it will form a sort of combined image of a statistically average cat, then use that to recognise further cats.
Deductions however are very specific and unique, so neural nets can't find anything statistically average about them to work with. Similarly, people don't have statistical conversations, they use language to describe unique objects and specific situations. The two words "my mother" mean an entirely different person from when you say "my mother". To a neural net however, it's just the same sequence of letters. Whenever you try to use neural nets on language, you get average answers rather than a specific one.

Google does a lot of great work, but language understanding is one area which I hope will ultimately be dominated by Good Old-Fashioned AI.

[ivan.moony]

Google offers an AI assistant, which is quasi-AI. There is one scenario I'm afraid of with AI: as the time pass, quasi-AI will get better and better and might even saturate low-level needs of the population. In that case, when someone creates true AI, he will find himself in very unpleasant situation where he wouldn't be noticed with his achievement.

But let's hope that true AI would start to make significant achievements in science, which might pave the road for the lucky creator.

[Art]

Google is so large, as is their language processing database that it can understand scores of data / pattern matching words and phrases by "brute force" alone. Examples of this were used in early chess programs where instead of relying upon skill and cunning, the program used a huge database of similar positions, exchanges and patterns to give itself an advantage which proved to be rather successful for its time.

Google Now, is still extremely adept at "understanding" human speech, even in some less-than-prefect conditions / environments. Give them a few more years and all this will no longer matter. Only the outcome really matters anyway.