The first person to hack the iPhone is building a self-driving car

Re:The first person to hack the iPhone is building a self-dr... by bloaf on Wednesday December 16, 2015 @12:35PM (#10745799)

I don't doubt it at all. Keep in mind that "simple" is relative, we have to ask "simple compared to what?" For lots of people, I be that the math involved in these neural nets is the most complicated math they've ever done. They would never say it is simple, because they themselves barely grasp it. But in my experience, topics in mathematics have a funny way of becoming very simple the moment you "graduate" to thinking about a slightly more general mathematical framework.

Someone who has digested enough of the AI literature to think about the methods in aggregate is very likely to be in a position to see any particular method as a "simple" implementation of some more general set of principles.

[ Parent | on HN ]

Re:The first person to hack the iPhone is building a self-dr... by mturmon on Wednesday December 16, 2015 @01:23PM (#10746124)

As a general observation, what you say has some truth to it.

But the particular quote is referring to learning rates in autonomous robotics, especially visual classification in complex real-world scenes.

I have worked and published in ML since the early 1990s, was a program chair for the learning track at NIPS one year, participated in the same DARPA learning-to-drive program that Yann LeCun did, and don't consider the math behind "state-of-the-art papers" to be simple.

Just taking deep learning: there are a lot of tricks and recipes (e.g., rectified-linear activations, number of layers, staged training) that are not mathematically understood. It's exciting, but mathematically still a jungle. Just because a neophyte can code and optimize a network does not mean that the math that explains why it actually works is simple. As engineers, we need to understand why it works before using it in a safety-critical situation.

[ Parent | on HN ]

Re:The first person to hack the iPhone is building a self-dr... by emcq on Wednesday December 16, 2015 @01:32PM (#10746177)

While a good point that simple is relative, if you specifically look at deep neural networks we don't understand how training a non convex function converges with gradient descent - the fundamental component to create a usable model. In practice, it often works, and there are a few intuitions why this works. But its naive at best to claim that this is simple. If it was we would understand it better :)

[ Parent | on HN ]