No academic endeavor has been more disappointing for me than Machine Learning. At some point I realized it was just statistics, it was prediction, it was regression... So why not just call it that? Where's the artificial "intelligence", where's the "learning"? The only intelligence to it is from the scientist who does the data mining, does the feature selection, and chooses the problem category.
Nah, that's the mentality that rebrands anything we actually figure out how to do as "not AI". It's an endlessly moving goalpost.
In 1956 a large Prolog program from today would have been considered clearly AI, but now it's "just logic programming". I don't think using the term AI for either logic programming or regression is inappropriate.
There are major differences between machine learning, especially deep learning, and traditional statistics. In particular, the focus on learning nonlinear hierarchical representations from raw data has enabled deep learning to make major advances in several applications beyond what traditional statistics could do.
A major benefit of Deep Learning is that it can learn without manual feature selection (i.e., Representation Learning).
To your other point, DeepMind's Atari-playing neural networks figured out strategies that their programmers were not aware of. Another example is AlphaGo 2017's tactics used in its self-plays that perplexed master Go players worldwide because they were not in any Go books or records in history but they are hugely effective and the pros are now learning from them.
I went through a similar experience, and spent two years going through the machine learning landscape. Every month or so I would wonder why I still bothered to continue working on it, but still continued as I saw so many smart people around me apparently interested in the topic.
It took a while before I realized that maybe ~10% of the researchers are in it out of genuine interest, and the remainder are simply hoping to ride the wave of success of various kinds (academic, monetary, etc).
And for the 10% who are genuinely interested in the stuff, there is a long history, with many ideas that got explored under various terms as you point out. It is extremely dangerous for a field to lose awareness of the history of ideas affecting it. This is part of what I like about older fields, such as mathematics: researchers at the highest level are very keenly aware of the history of ideas in the discipline, and where modern ``high powered machinery'' can and can't help one on various problems. One can only appreciate the problems and techniques of today's enterprise if one understands what came before.
An analogy I like to think of is that of the problem of alchemy in the middle ages. Turning base substances into gold was a hopeless task back then (and still is right now); yet very intelligent people of the time such as Newton devoted significant amounts of energy into it. From a modern scientific perspective, there is nothing fundamentally impossible about this; stars do it in supernovae. What the modern perspective has given us is the ability to see why the problem is out of reach for now, and why energies are better devoted elsewhere.
My own belief is that 100 years from now, people will be able to articulate similar reasons with respect to the problem of strong AI. A personal favorite in terms of arguments are those of Stanislaw Ulam, as presented in Chapter 4 of "Indiscrete Thoughts" by Gian-Carlo Rota.
In the meantime, there are side benefits to be had from the pursuit of the ``strong AI challenge''; in a similar manner to the great progress in chemistry thanks to the problem of alchemy in the middle ages. Grand challenges serve as a unifying cause, and thus help in achieving remarkable things.
7 comments
[ 2.6 ms ] story [ 30.4 ms ] threadIn 1956 a large Prolog program from today would have been considered clearly AI, but now it's "just logic programming". I don't think using the term AI for either logic programming or regression is inappropriate.
A major benefit of Deep Learning is that it can learn without manual feature selection (i.e., Representation Learning).
Here's a good article: http://www.sciencedirect.com/science/article/pii/S2405918816...
To your other point, DeepMind's Atari-playing neural networks figured out strategies that their programmers were not aware of. Another example is AlphaGo 2017's tactics used in its self-plays that perplexed master Go players worldwide because they were not in any Go books or records in history but they are hugely effective and the pros are now learning from them.
It took a while before I realized that maybe ~10% of the researchers are in it out of genuine interest, and the remainder are simply hoping to ride the wave of success of various kinds (academic, monetary, etc).
And for the 10% who are genuinely interested in the stuff, there is a long history, with many ideas that got explored under various terms as you point out. It is extremely dangerous for a field to lose awareness of the history of ideas affecting it. This is part of what I like about older fields, such as mathematics: researchers at the highest level are very keenly aware of the history of ideas in the discipline, and where modern ``high powered machinery'' can and can't help one on various problems. One can only appreciate the problems and techniques of today's enterprise if one understands what came before.
An analogy I like to think of is that of the problem of alchemy in the middle ages. Turning base substances into gold was a hopeless task back then (and still is right now); yet very intelligent people of the time such as Newton devoted significant amounts of energy into it. From a modern scientific perspective, there is nothing fundamentally impossible about this; stars do it in supernovae. What the modern perspective has given us is the ability to see why the problem is out of reach for now, and why energies are better devoted elsewhere.
My own belief is that 100 years from now, people will be able to articulate similar reasons with respect to the problem of strong AI. A personal favorite in terms of arguments are those of Stanislaw Ulam, as presented in Chapter 4 of "Indiscrete Thoughts" by Gian-Carlo Rota.
In the meantime, there are side benefits to be had from the pursuit of the ``strong AI challenge''; in a similar manner to the great progress in chemistry thanks to the problem of alchemy in the middle ages. Grand challenges serve as a unifying cause, and thus help in achieving remarkable things.