Thanks for the upvotes! I'm a little conflicted about this being linked around too much because it is still very much work in progress. I work on this guide on a side because I think there's a lot of interest in these models out there, and not enough explanations.
On a related note, some of you might also be interested in a Stanford CS class I'll be co-teaching with my adviser next quarter. It will be focused primarily on Convolutional Networks (but a lot of the learning machinery is generic):
I hope to make a lot of the materials/code freely available so everyone can follow along, and I will continue my work on this guide in parallel whenever I can squeeze in time. (And I'd be happy to hear any feedback on the content/style/presentation)
1) In "Becoming a Backprop Ninja", dx is never declared. Is it a variable that was set in a previous pass, or is it some value like 1 that depends on its function? I understand how to derive dx1 and dx2 from dx.
2) In the SVM, can "pull" be other values than -1, 0, 1? It seems like it might affect how rapidly the SVM learns.
3) It takes 300 iterations to train the binary classifier, could you add a brief paragraph about why this isn't 30 or 3000, and what has the greatest effect on iterations?
4) In 2D SVM, what is the α function? Are w0^2 and w1^2 an arbitrary way to regularize (like would |w0| + |w1| work as well?) It made me think of least squares fitting or RMS power, so I wondered what the reasoning was behind it.
Many more questions (and a sense of irony because I have learned NNs in the past and have since forgotten) so I am trying to give my brain some breadcrumbs.
Thank you for publicizing an early draft. It is really good and very helpful.
I think your explanations are great, but you seem to imply that the gradient 3 for f(x,y) = xy with x = -2 and y=3 derived w.r.t x is somehow the "best" gradient (you call it the correct or actual gradient) but I don't understand why this is better than any other positive number.
Wouldn't we optimize faster of we tug with a force of 6 instead of 3?
x = -2, y= 3,
dx = 3, dy = -2
step = 0.0001
f(x,y) = -6
f(x+dx*step, y*dy*step) = -5.998...
// Now if we double the gradient
f(x+2*dx*step, y*2*dy*step) = -5.997...
The analytical gradient is actually a general tool used in multivariable calculus. It's vector-valued, in that if you take the gradient of a function of three variables: f(x, y, z), then you'll get a 3-vector back. Vectors are defined by two characteristics: a direction and a magnitude. The gradient's direction is the direction of greatest increase and the magnitude of the gradient is the instantaneous rate of change.
The gradient is being put to work here in order to optimize a function using a process called gradient ascent. Intuitively it makes sense that in order to optimize your function you'd want to "move" in the direction of greatest increase at every step. However, the size of the step that you take is tricky. As you point out, in this case, we can increase the objective function value more if we double the step size. However, you're not actually doubling /the gradient/.
If you look at the expression that you wrote you should see:
x + 2 * dx * step, y * 2 * dy * step.
What you've done is double the step-multiplier, not the gradient itself (<dx, dy>). This means that the optimization process jumps further at each step. However, the step size that OP chose is somewhat arbitrary to begin with, so it's not clear why any particular choice would be better or worse. The reason that the step size matters is that if your step size is too large your optimization process might jump over a minimum location or something similarly bad -- there are pathological functions like the Rosenbrock Function [1] which are notoriously hard to optimize with gradient ascent. In practice, you'll often choose your step size more intelligently based on a few different tests, or vary it as the optimization process progresses.
In this particular instance, the surface that you're trying to optimize is pretty simple so basically any step value will do the trick. It may take a different number of steps in order to compute the global maximum, but most reasonable step sizes will get there.
This is actually a Really good question, I should really expand on this in the guide.
The simplicity of the example is deceiving you, in the sense that in a more complex circuit (than just x * y), it will no longer be the case that just simply marching along the direction of the gradient in huge steps will necessarily lead to good outcome.
All we can do is work very locally. You are at some specific point (x,y)=(-2,3). The gradient is telling you that the best direction (in this 2-D space) to go along if you're interested in increasing your function is in the direction (3, -2), which is the gradient at that point.
Now if you remember there's a crucial parameter step_size, and the update will take the form (in vector notation):
(x,y) += step_size * (dx, dy)
all that the gradient is saying is that if step_size is infinitesimally small, your function will increase, and that this direction is the Fastest increase. If you try any other direction with infinitesimally small step_size, you will do worse. However, there are no guarantees whatsoever on what happens when the step_size is larger. In practice we use step sizes of 0.001, or whatever (small numbers), and it works only because the functions we deal with are relatively smooth and well-behaved.
However, after we take the small step we have to right away re-evaluate the gradient because suddenly the direction could be all shifted. Sometimes doing gradient descent is compared to walking blind on a hill and trying to get to the top: you can sense the steepness of the hill at your feet (the gradient), and you make steps accordingly. But if you make too large of a step and you're not careful, there could have been actually been large drop.
TLDR: x * y is very simple example, this certainly wouldn't be the case in more complicated circuit.
I'm a huge fan of interactive demos (see e.g. my ConvNetJS demos: http://cs.stanford.edu/people/karpathy/convnetjs/ ) because they help develop intuitions that solidify the material. And yes, that's partly why I chose Javascript; I'm definitely planning to add little demos of the gradients flowing around the circuit and so on. Coming soon! :)
On a related note, some of you might also be interested in a Stanford CS class I'll be co-teaching with my adviser next quarter. It will be focused primarily on Convolutional Networks (but a lot of the learning machinery is generic):
CS231n: Convolutional Neural Networks for Visual Recognition http://vision.stanford.edu/teaching/cs231n/
I hope to make a lot of the materials/code freely available so everyone can follow along, and I will continue my work on this guide in parallel whenever I can squeeze in time. (And I'd be happy to hear any feedback on the content/style/presentation)