Andrew Ng on What's Next in Deep Learning [video]

His position seems to be similar to MIT engineer David Mindell's, who recently wrote Our Robots, Ourselves[1]. It seems to me that the more hours one spends solving real problems with AI, the less likely they are to believe an apocalypse is imminent. Conversely, the more dependent a prognosticator is on book or film sales, the more likely they are to predict or depict a near-term AI apocalypse.

[1] Google Talk: https://www.youtube.com/watch?v=4nDdqGUMdAY

He may not have meant to imply this, but I think he's wrong on at least one of the points.

He says there's no way to work productively on the problem. And perhaps from one perspective he's right, there is no imminent threat of overpopulation on Mars, and so probably no one is going to pay him a salary to work on it, but in general I think preventative maintenance is more important than dealing with fires when they actually happen.

If we know "overpopulation" is a problem in modern civilization, why not work on defining measures to prevent it in the future instead of letting history repeat itself?

And as far as AI apocalypse is concerned, I'm not concerned from one perspective but somewhat concerned from another. I think we don't need to be worried that AI will become sentient in the near future. I think we do need to be concerned about humans putting too much trust in AI to solve / automate critical problems for us. As a perfect example, I wouldn't recommend putting the decision of launching a nuclear weapon in the hands of some AI system. Though the AI wouldn't be evil in its intention to launch nuclear bombs inadvertently, the inadequacies of its ability to understand when it would be appropriate to launch nuclear bombs would certainly cause a lot of problems the world would need to deal with in the aftermath (if anyone were to survive).

Andrew Ng's course on Coursera?

It is not necessary to use phonemes in the model, however, it greatly restricts model's ability to learn, in particular to learn many special cases like rare foreign words. You need much more data to train the model and basically you lose ability to quickly teach the system a new word by specifying word pronunciation. It might sound great for advertising of computing power factories like Baidu is doing, but speech recognition experts are not that enthusiastic about such approach.

It is pretty disappointing that modern machine learning "experts" advertise their improved results with questionable comparisons and do not care about machine learning theory, in particular, never consider model ability to generalize, model robustness to noises and related things which must be primary subjects for research.