An intuition of Newton's method

During my lazy weekend afternoons (and all the other days of the week) I've been going through Nando de Freitas' undergraduate machine learning course on youtube. In lecture 26 he introduces gradient descent, an iterative algorithm for optimizing (potentially non-convex) functions. I'll refrain from explaining gradient descent more than that as there are many good explanations on the internet (including Professor de Freitas'), but I do want to discuss gradient descent's learning rate parameter and why intuitively Newton's method, also introduced in lecture 26, is able to side-step the learning rate.

As described in the video lecture, a common difficulty with gradient descent is picking the learning rate hyper-parameter. Pick too small of a learning rate and gradient descent will come to crawl and make very little progress or even eventually stop making progress due to floating point underflow. Pick too large of a learning rate and gradient descent will oscillate around the minimum/maximum, bouncing around, very slowly making progress. In an ideal world, as the gradient vanishes, the gradient descent algorithm would be able to compensate by adjusting the learning rate so to avoid both underflow and oscillation. This is exactly what Newton's method does.

Compare gradient descent
$$\boldsymbol\theta _{k+1} = \boldsymbol\theta _{k} - \eta_{k}\nabla f(\boldsymbol\theta_{k})$$
to Newton's method
$$\boldsymbol\theta _{k+1} = \boldsymbol\theta _{k} - H_{k}^{-1}\nabla f(\boldsymbol\theta_{k})$$

The only difference is that Newton's method has replaced the learning rate with the inverse of the Hessian matrix. The lecture derived Newton's method by showing that \(f(\boldsymbol\theta)\) can be approximated by a second-order Taylor series expansion around \(\boldsymbol\theta_{k}\). If you're familiar with Taylor series, then this explanation may be sufficient, but I prefer to think about it differently.

What is this Hessian matrix we've replace the learning rate with? Just as the gradient vector is the multi-dimensional equivalent of the derivative, the Hessian matrix is the multi-dimensional equivalent of the second-derivative. If \(\boldsymbol\theta_{k}\) is our position on the surface of the function we're optimizing at time step \(k\), then the gradient vector, \(\nabla f(\boldsymbol\theta_{k})\), is our velocity at which we're descending towards the optimum, and the Hessian matrix, \(H\), is our acceleration.

Let's think back to the original problem of choosing the learning rate hyper-parameter. It is often too small, causing underflow, or it is too big, causing oscillation, and in an ideal world we could pick just the right the learning rate at any time step, most likely in relation to the rate at which the gradient vector vanishes, to avoid both these problems. But wait, now we have this thing called the Hessian matrix that measures the rate at which the gradient vanishes!

When starting the algorithm's descent we will quickly gain acceleration as we starting going downhill along the function's surface (i.e. the Hessian matrix will "increase"). We don't want to gain too much velocity else we'll overshoot the minimum and start oscillating, so we'd want to choose a smaller learning rate. As the algorithm nears the optimum, the geometry of the function will flatten out and we will eventually lose acceleration. We don't want to lose too much velocity else we'll underflow and stop making progress. It is precisely when we lose acceleration (i.e. the Hessian matrix "decreases") that we want to increase our learning parameter.

So we have observed that heir is an inverse relationship between the magnitude of the acceleration at which we descend towards the minimum and the magnitude of the most desirable learning parameter. That is to say, at greater accelerations we want smaller learning rates and at smaller accelerations we want larger learning rates. It is precisely for this reason why Newton's method multiplies the gradient vector (the "velocity") by the inverse Hessian matrix ("the inverse acceleration")! At least, this is how I intuitively think about Newton's method. The derivation involving Taylor series expansions is probably the actual precise reason, but I like thinking about it in terms of velocities and accelerations.

Finally, I'd like to end with one cautionary note. I'm an amateur mathematician. I like to think my intuition is correct, but it may not be. By sharing my intuition, I hope to help others who are looking to gain intuition, but if you need anything more than an intuitional understanding of Newton's method or gradient descent (because perhaps you're an actual mathematician in training), please consult a textbook or your favourite professor :-).

On optimizing high-dimensional non-convex functions

Excuse me for the (in)completeness of this post. What follows is merely a thought, inspired by two independent statements, about a domain of science (or math, really) with which I am barely initiated. Let me give you these two statements first.

In the book Foundations of Data Science, the very first paragraph of the book says

If one generates n points at random in the unit d-dimensional ball, for sufficiently large d, with high probability the distances between all pairs of points will be essentially the same. Also the volume of the unit ball in d-dimensions goes to zero as the dimension goes to infinity. 

That is statement one. Statement two is an assertion given in this year's Deep Learning Tutorial at the Neural Information Processing Systems (NIPS) conference. They claim that "most local minima are close to the bottom (global minimum error)."

The second statement is significant for training (deep) neural networks. Training a neural network always uses a variant of a gradient descent, a method for minimizing the neural network's error by iteratively solving for when the gradient (i.e. the derivative) of the function that gives the error is 0. Ideally gradient descent will approach the global minimum error (the point where the neural network will give the most desirable approximations of the training data), but that is only guaranteed when the error function is convex (there are no hills when the function is plotted). Unfortunately, a neural network almost never has a convex error function. Hence, during gradient descent, one might find a local minimum that is nowhere near as low as the global minimum. The second statement however says we shouldn't be so concerned -- all local minima will be close to the global minimum!

But why?

My thought is that somehow the first statement can be used to prove the second statement. The training data of modern neural networks is high dimensional. It will also often be normalized as part of feature scaling. Given those two properties, the high-dimensional hills and valleys of a neural network's error function then (roughly) form unit hemispheres. This (possibly) implies that that local minima are close to the global minimum because the volumes of every valley nears zero, making all the valleys increasing similar to each other as the dimensionality of the training data increases.

I want to stress though, this is a half-baked idea. It's probably rubbish. It might also be redundant! Perhaps the second statement already has a proof and I've missed it. Either way, I would love to see if others find my intuition plausible or, even better, to be pointed in the direction of a proof for the second statement.