It's interesting that they say it's been applied to NLP but provide no description on results, which makes one think it didn't do as well as a NL classifier as it did for the image classifying application. I wonder if that's a result of training size, or there's something more inherent in undering NL that just makes it much harder, and that DNNs won't do as well on in comparison to Image classification? That later point is interesting for AGI, because it implies that there's still a larger ML paradigm that's needed for NL, before we'll see real breakthroughs.
They also talk about these big improvements in accuracy, but give no figures. It'd be nice to know what it's relative to. Guess we have to wait for the paper... -Chris On Tuesday, July 15, 2014 6:07 AM, Jim Bromer via AGI <[email protected]> wrote: It is scalable they say. OK, but in how many cognitive categories? Are the neural networks able to use that information for combination and selection at a higher cognitive order? I am asking if they are able to use their capabilities in a creative and imaginative application of knowledge. If you could get a neural network to detect imagination then you might be able to train it to use directed imagination in problem solving. But you can't because imagination is too broad a concept. Jim Bromer On Tue, Jul 15, 2014 at 5:51 AM, Jim Bromer <[email protected]> wrote: It is scalable they say. OK, but in how many cognitive dimensions? (Throwing a term like "cognitive dimensions" out in front of you guys may be a mistake.) Perhaps I should have asked, in how many cognitive categories? This has been a persistent problem with neural networks. The point of view that Chilimbi seems to be taking is that the [parallel] system is able to detect different features and combine them somehow. However, that does not mean, for example, that the neural networks are able to use that information for combination and selection at a higher cognitive order. Just because I cannot word my criticism in just the right way does not mean that the criticism is completely invalid. A counter argument is that since the algorithms are effectively making some kinds of combinations and some kinds of selections then that proves that they are theoretically capable of making the kinds of subsequent combinations and selections that I am thinking of. However, this (kind of) argument still hasn't been established experimentally. It's the same old same old. > > >Jim Bromer > > >On Tue, Jul 15, 2014 at 12:22 AM, Ben Goertzel via AGI <[email protected]> >wrote: > > >>;) >> >>Hmm.. the quote about quantum physics is >> >>*** >>“It’s like in quantum physics at the beginning of the 20th century,” Chilimbi says. “The experimentalists and practitioners were ahead of the theoreticians. They couldn’t explain the results. We appear to be at a similar stage with DNNs. We’re realizing the power and the capabilities, but we still don’t understand the fundamentals of exactly how they work. " >>*** >> >>It is true that with deep learning, practice is way ahead of theory these >>days. OTOH, quantum mechanics has some mysterious confusing-ness too it, >>which deep learning doesn't really have.... What modern DL algorithms are >>doing has made sense conceptually to a lot of people for a long time, and the >>recent breakthrough results are not based so much on new conceptual >>breakthroughs, as on the new availability of massive amounts of hardware and >>associated operational software.... >> >>Still, "solving object recognition via DL" is a fair bit more AGI-ish than >>"solving chess via alpha-beta pruning", for example These DL architectures >>have a strong conceptual resemblance to one aspect of the human mind-brain >>(parts of our visual and auditory cortices), whereas alpha-beta pruning has a >>significantly lesser resemblance to stuff that the human mind does (the human >>mind does some search like that, but on a far lesser scale and only in tight >>interaction with a lot of other processes).... >> >> I think these modern DL architectures bear moderately close resemblance to >>something that could serve as a significant component of an AGI system. In >>their details, though, they are not architected for integration with other >>AGI components, unlike how the human visual and auditory cortices, with their >>pronounced hierarchical sensory pattern recognition functionality, are >>architected (well, evolved) for interaction with the rest of the brain... >> >>-- Ben >> >> >> >> >>On Mon, Jul 14, 2014 at 9:04 PM, David Hart <[email protected]> wrote: >> >>I skimmed it. Is it a joke? DNN like quantum physics? Really? Do they >>>really think the two are in the same league? >>> >>>-dave >>> >>> >>> >>>On Tue, Jul 15, 2014 at 11:52 AM, Ben Goertzel <[email protected]> wrote: >>>> >>>> This time from Microsoft... >>>> >>>> http://research.microsoft.com/en-us/news/features/dnnvision-071414.aspx >>>> >>>> -- >>>> Ben Goertzel, PhD >>>> http://goertzel.org >>>> >>>> "In an insane world, the sane man must appear to be insane". -- Capt. James >>>> T. Kirk >>>> >>>> "Emancipate yourself from mental slavery / None but ourselves can free our >>>> minds" -- Robert Nesta Marley >>> >>> >>> >>>-- >>>David Hart | +852 9559 9580 >>>OpenCog Foundation | http://opencog.org >>> >> >> >> >>-- >>Ben Goertzel, PhD >>http://goertzel.org >> >>"In an insane world, the sane man must appear to be insane". -- Capt. James >>T. Kirk >> >>"Emancipate yourself from mental slavery / None but ourselves can free our >>minds" -- Robert Nesta Marley >>AGI | Archives | Modify Your Subscription > AGI | Archives | Modify Your Subscription ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
