Artificial Intelligence: Modern Magic or Dangerous Future? by Yorick Wilks

Author:Yorick Wilks [Wilks, Yorick]
Language: eng
Format: epub
ISBN: 9781785785177
Google: 966WDwAAQBAJ
Amazon: 1785785168
Goodreads: 42944390
Publisher: Icon Books
Published: 2019-06-05T23:00:00+00:00

All the methods in the list above had effects on NLP and AI generally; all made very large-scale claims that the NLP problem was essentially solved; none produced the results they promised but none of them ever entirely went away. Jelinek’s methods completely redirected the course of NLP from the mid-1990s onwards, even though it was the technique least likely to model human translation, because of the vast exposure to translated texts it required. All of them hit barriers that prevented large applications in their original forms: early Connectionism could not extend beyond nets of trivial size; Jelinek’s MT system never really beat the SYSTRAN hand-coded MT system in open competition, and reinforcement learning systems failed to learn anything very interesting.

But it must be said that the limits to deep learning are not yet known. The distinguishing feature of DL, apart from having more layers of nodes, was intended to be that it did not require features on which to learn to be given to it, but that such systems could discover them for themselves, and be truly unsupervised, to use the older term we defined earlier. But it has not proved so simple to eliminate all such promptings of systems, and the fundamental discussions in DL are now about whether that is actually possible. The mystery of human learning means that DL programs often fail to learn from millions of examples, although humans seem to learn new concepts from two or three – as we saw with steampunk dress – but we have no idea how they do it.

This may suggest that written text understanding presents challenges of a kind that the techniques we have (ML and non-ML) do not yet solve, nor are likely to in the immediate future. And this is perhaps because written language has aspects of a crafted artefact, rather than a function naturally developed under evolutionary pressures such as vision and speech. Saying this is not fashionable, not least because it goes against the prejudice of most linguists that speech is primary and is real language, and its written form is secondary. The nature of the speech–language difference is harder to see from inside a language such as English where the written form can be converted fairly straightforwardly to a spoken form by the ASR technology we described earlier. But from inside a language such as Chinese, the issue looks completely different, since a written string can be spoken in a number of mutually incomprehensible ways, some of which can even be in a language from a different language family, such as Japanese! Yet understanding written text automatically is perhaps the primary goal of current AI technology.

Two other aspects of the modern ML systems mentioned above should be brought out. The key notion of distributed representations goes back to Connectionism and is one of the features marking off modern ML from the older, symbolic rule-driven forms of ML. This is the view that what we know about the brain and evolution does not support

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.