Has Google's AI invented it's own language?
We’ve had a number of discussions about whether computers can be creative or if they just do “what they were programmed to do”. Google Translate revamped their algorithms last September. Previously, they had programmed the system to have phrase pairs between different languages. But with 103 languages the number of pairs they needed to have was growing beyond their ability to maintain it.
They switched from a programmed translation to a neural network and fed the vast database they already had it. The resulting translation utility is now live and now appears to have developed an internal representation language that it translates source languages to and then to the destination language. It was not specifically programmed to do this.
This allows it to translate between pairs of languages that it doesn’t have an example for. A human who knew the French to English translation of a phrase and the English to Russian one, might translate French to Russian by going through English. The Google translation translates through the language it made up.
They switched from a programmed translation to a neural network and fed the vast database they already had it. The resulting translation utility is now live and now appears to have developed an internal representation language that it translates source languages to and then to the destination language. It was not specifically programmed to do this.
This allows it to translate between pairs of languages that it doesn’t have an example for. A human who knew the French to English translation of a phrase and the English to Russian one, might translate French to Russian by going through English. The Google translation translates through the language it made up.
Jan
If there is a composite language created from all of the languages extant, then it might actually be better at expressing all of the concepts we need, and it will probably be more logically constructed. Whether or not it is suitable for humans to use is, as you point out, another question entirely.
Jan
I imagine Univac to Univac was written in this language, and I read an English translation.
The google report is interesting but does not show consciousness, focus or creativity by an independent 'mind'. It is the result of how the neural net was programmed and the data fed into it. The internal state of a neural net evolves over time as the data accumulates, in accordance with whatever is defined as 'success'. The evolved state of net is never directly programmed.
In this case the internal representation that evolved converged on the essence of the language structure common to all the languages. It is interesting how it wound up with a state so efficient and that it worked at all, but if it does work at all it had to be some representation of language -- and the net had no means of distinguishing between languages other than what it was fed to help classify but not completely restrict. So in a crude sense, the different languages are all ways of saying the same thing just like there are different ways within a language in both grammar and vocabulary.
The representation seems to have succeeded where the direct attempt by man did not: univacs couldn't understand Esperanto (which probably contributed to their fears of man).
http://www.research.ibm.com/cognitive...