Ravi, (01)
> I ... agree that computers and ANN and AI are very far from
> human intelligence... (02)
Yes. But the AI field has been plagued by exaggerated hype for the
past half century. I agree with Gary Marcus, a psycholinguist at NYU
who has also worked with the Allen Institute (which is doing important
research in AI):
http://www.newyorker.com/tech/elements/hyping-artificial-intelligence-yet-again (03)
A reader who sent me an offline email, suggested the following book
by Le Deng and Dong Yu, two researchers at Microsoft:
http://research.microsoft.com/pubs/209355/DeepLearning-NowPublishing-Vol7-SIG-039.pdf (04)
I've only had a chance to browse through the book and read selected
parts. But it's a well-written, hype-free, 190-page presentation of
the history, methods, and applications of DNNs (where D stands for
'deep'). The authors admit that DNNs do *not* support a human level
of "deep understanding". They quote a definition by the educational
psychologist Eileen Hulme (2009): (05)
From
http://www.blackwellreference.com/public/tocnode?id=g9781405161251_chunk_g97814051612516_ss1-1
> Deep learning describes an approach to learning that is characterized
> by active engagement, intrinsic motivation, and a personal search
> for meaning. Deep learners actively seek a significant conceptual
> understanding by connecting new concepts with previously held
> knowledge. They rigorously evaluate evidence, examine logic, and
> critically analyze hypotheses and conclusions. (06)
On pp. 19 to 21 (by the counter in the Adobe reader), Deng & Yu
discuss the Gartner "hype cycle". On p. 21, they show the graph
of "Expectations or media hype" over time. The "peak of inflated
expectations" occurred in 1990, the "trough of disillusionment" in
2006", and we're now in a "plateau of productivity" -- but they mark
that plateau with "?". (07)
I think that's a fair assessment. If all the DNN advocates stated
those qualifications, I'd have no quarrel with them. (08)
On pp. 24 to 39, they distinguish the three major kinds of DNNs with
many citations, including comparisons with related methods. (09)
I browsed through the remaining chapters and paused to read various
parts. They are clearly written, and they cite other publications
for details. The following sentence on p. 141 captures the essence
of what DNNs do that distinguishes them from the earlier ANNs: (010)
> The essence of deep learning is to automate the process of discovering
> effective features or representations for any machine learning task,
> including automatically transferring knowledge from one task to
> another concurrently. (011)
The ability to *discover* features automatically is critical. Experts
in any field usually assume that the terminology of the field covers
the most important aspects. But even the experts are unable to state
which combinations of terms are the most relevant. Various statistical
methods (including ANNs and DNNs) can often discover combinations that
even the experts overlooked. (012)
However, note the definition by Eileen Hulme above. Human "deep
learners" do a lot more than discover combinations. That's the point
of the cognitive cycle that I emphasized in slides 41 to 57 of
http://www.jfsowa.com/talks/micai.pdf (013)
Note slide 52 of micai.pdf: it shows a variety of different AI
methods that have proved to be useful for the steps in the cycle.
The ones I listed are just a small sample of the open-ended variety
of AI methods that have been used and new ones that may be invented. (014)
Also note Ohlsson's "deep learning cycle" in slide 46. I recommend
his book and its subtitle "How the mind overrides experience". By
his classification, DNNs are primarily pattern recognizers. But the
"mind" can override recognition and suggest other interpretations by
shifting attention. See the reference to "active learning" in slide 54. (015)
Also note slide 55: "Human learning requires language." Sometimes
language can mislead or bias the interpretation. But all mammals can
do the kind of pattern recognition that DNNs do, and they aren't as
smart as humans -- because they don't have language. (016)
Without language, the human kind of deep learning (as Ohlsson or Hulme
define it) is impossible. Nothing in the book by Deng & Yu begins to
address the issues that Ohlsson discusses or Hulme mentions. (017)
John (018)
_________________________________________________________________
Message Archives: http://ontolog.cim3.net/forum/ontolog-forum/
Config Subscr: http://ontolog.cim3.net/mailman/listinfo/ontolog-forum/
Unsubscribe: mailto:ontolog-forum-leave@xxxxxxxxxxxxxxxx
Shared Files: http://ontolog.cim3.net/file/
Community Wiki: http://ontolog.cim3.net/wiki/
To join: http://ontolog.cim3.net/cgi-bin/wiki.pl?WikiHomePage#nid1J (019)
|