Let's say you're building a robot tennis player. Ignoring the physical engineering for a second, let's assume it can move around and swing a racket with roughly the same speed as a human.
It doesn't have anything other than a camera to track the ball.
So in terms of it's control, we have two options:
1: track the ball in 3d, work out the velocity, trajectory, ballistics, etc and then calculate the optimal position and swing to return.
2: use some form of ML that recognises patterns and applies patterned responses.
I'm willing to bet that 2 will significantly outperform 1, for a fraction of the compute power (once the training data is done).
Because that's what humans do. No one is working out equations in real time.
Human brains aren't bad at this, but I'd bet large sums of money that machines will out perform us at this within a decade.
> Human brains aren't bad at this, but I'd bet large sums of money that machines will out perform us at this within a decade.
I'm not so sure. Many other spaces where computers outperform us come nowhere close to doing it in the same power budget, and the amount of power it takes isn't reducing quickly.
> Because that's what humans do. No one is working out equations in real time.
I believe we do something in-between-- like have a kinematics and tracking engine that is fuzzy and gets tuned/updated/supplemental training for particular problems we face a lot, but isn't very tennis-motion specific.
I’m not sure human brains don’t do equations in real time. Sure, not in the form a computer would do it, nor in a way we could use up intermediate calculations, but humans have great spacial/visual skills which are not solely akin to ML algorithms. I’m a layperson, but what I mean here is rotating a 3D object in your head is not the same mechanism as your general flow of thought. It can perhaps be thought of as “graphic acceleration”.
I think both brains and ML in practice "solve equations". The main thrust of machine learning is using linear algebra as a universal function fitter. Sure, the state space may be transformed in interesting ways and the overall process may be relatively opaque, but you could easily end up with the quadratic formula in there.
The equations may be very strange from the point of view of the underlying physics/kinematics problem. There's a lot of representations of the problem that fit well and may map better to what our perceptions see and the level of precision that our muscles need.
And it's likely very messy--- we don't have some clean, nested set of controllers here, but strange ways that many kinds of feedback are combined.
I honestly doubt ML will outperform a classical approach. A tennisbal is relatively easily recognized using basic filters. The kinematic equations to solve are cheap.
It doesn't have anything other than a camera to track the ball.
So in terms of it's control, we have two options: 1: track the ball in 3d, work out the velocity, trajectory, ballistics, etc and then calculate the optimal position and swing to return. 2: use some form of ML that recognises patterns and applies patterned responses.
I'm willing to bet that 2 will significantly outperform 1, for a fraction of the compute power (once the training data is done).
Because that's what humans do. No one is working out equations in real time.
Human brains aren't bad at this, but I'd bet large sums of money that machines will out perform us at this within a decade.