4 ms·
Term 'Artificial Intelligence' is vague. "Deep learning is a name for machine learning techniques using many-layered artificial neural networks. Occasionally p
by stared 9y ago
Term 'Artificial Intelligence' is vague.
"Deep learning is a name for machine learning techniques using many-layered artificial neural networks. Occasionally people use the term artificial intelligence, but unless you want to sound sci-fi, it is reserved for problems that are currently considered “too hard for machines” - a frontier that keeps moving rapidly." from http://p.migdal.pl/2017/04/30/teaching-deep-learning.html http://p.migdal.pl/2017/04/30/teaching-deep-learning.html
In particular, two decades ago, letter recognition was considered a holy grail of AI, see:
"Indeed, I once even proposed that the toughest challenge facing AI workers is to answer the question: “What are the letters ‘A’ and ‘I’?" - Douglas R. Hofstadter (1995)
Not it as simple assignment to anyone learning ConvNets (BTW: I really recommend notMNIST as a starting point).
- ska 9y agoI've seen that Hofstadter quote used in this context before, and always assumed it was being quoted out of context. Do you know where it comes from? I suspect the idea is related to his "Letter Spirit" project - if so approaches like notMNIST don't really speak to it. Hofstadter often wrote about the difficulty in semantic understanding, and how AI approaches often approached things in such a superficial way that they miss the deeper meaning by design. Many of his projects were aimed at taking on these much harder problems head on. I suspect this quote, while also being a play on words, is aimed at this deeper understanding.
- stared 9y agoI give some context in the blog post linked above. This quotation comes from https://web.stanford.edu/group/SHR/4-2/text/hofstadter.html https://web.stanford.edu/group/SHR/4-2/text/hofstadter.html. And yes, he expected machine to have some 'deep understanding' of what is a letter. My approach (partly in the lines of Dennett, partly my experience with deep learning) is that we overestimate human cognition, which may be more 'a bunch of heuristics' than something 'really deep'. Especially, as some problems can be solved by... pigeons: http://www.sciencemag.org/news/2015/11/pigeons-spot-cancer-well-human-experts http://www.sciencemag.org/news/2015/11/pigeons-spot-cancer-w....
- ska 9y agoIn that context, it makes no sense to claim that convolutional NN have solved his problem at all, let alone claim it is a simple assignment. It is interesting to ask if he was asking the right question or not, but misleading to trivialize it.
- stared 9y agoWell, it is exactly what I meant this "artificial intelligence" is line that changes. First we consider something some problem too hard for machines. Then, once we solve it - it's "well, but it is just a singular case". In any case, what do you mean by "the right question"?