Visualizing the Math Behind Logistic Regression and Newton's Method

I wrote this post (first) on Gradient Descent that will be helpful if you liked the OP.

http://thelaziestprogrammer.com/sharrington/math-of-machine-...

This looks at the algorithm the same way that the OP does (mathematically, visually, programmatically), only it applies it to a simpler Linear Regression model.

In short, Newton's method uses second order derivative information in the search direction, while gradient descent only uses first order derivative. In between, there are "quasi-newton" methods which include generalizations of the "secant method". I should also mention that there are all sorts of ad-hoc approaches for attempting to increase the convergence rate of gradient descent, e.g., pre-conditioning, "momentum" terms, etc.

The calculus answer is that gradient descent approximates the shape of your optimization space with the first order taylor expansion at your currently optimized point (first derivative information), and Newton's method approximates it with the second order taylor expansion (first and second derivative information).

The geometric answer is that gradient descent approximates the shape of your currently optimized point by a plane that is tangent to the current point, whereas Newton's approximates it with a quadratic "bowl" using the local curvature. In gradient descent you have to take short steps down the path, because you know your approximation cannot be true (you will not always be able to lower your error by moving in the same direction, if you assume there is some minimum error). In Newton's method you can assume that your bowl approximation is correct and go straight to the bottom of that bowl. If your space really does look like a bowl, you will get to the bottom super quickly.

The ELI5 answer: you are a ship captain looking for the lowest point in a part of the ocean. You cannot afford to send a submersible down, so your approach is to send an anchor down to check the depth and take a panoramic photo a few feet around the anchor. After analyzing the results of each anchor drop you can move your ship anywhere and try again. In gradient descent you look for the lowest point on the edge of the camera's vision and move in that direction. You can move as far as you'd like, but you will not go too far because you know the drop will not go on forever. In Newton's method you will look at the curvature in the pano and extrapolate it out assuming it is a "bowl" shape. For example maybe your anchor is on an underwater J-shaped cliff getting less steep as it goes north, so you map it out as a U shape that starts going up again if you go a mile north. So, you move your ship a mile north and try again there. In gradient descent you would have moved maybe 100 feet north then checked again. Newton's method can intuitively get you to the low point in very few anchor readings, but crunching the numbers of the bowl can be very time-consuming (this becomes relevant in higher dimensions). In this analogy, the long and lat coordinates in the ocean are your two parameters to be optimized, and the height of the ocean floor is the error in your optimization. You want to find the parameters that give you the lowest error.

Thanks for the feedback, I co-worker just jabbed me with regarding the log property mistake also...

As to the motivation for the correct step: can you point me to a resource that explains this? Not sure I follow...

I'd also like to note that the Hessian matrix elements should have a \partial{\partial{l}} in each numerator, not a single \partial{l} [1].

If you aren't using LaTeX for formatting, then think partial^2 of l. FWIW: I just found this[2] which would make (using the physics package) even simpler to represent.

[1] https://en.wikipedia.org/wiki/Hessian_matrix

[2] https://tex.stackexchange.com/questions/225523/how-to-write-...

The motivation for the Hessian is the same as for dividing by the second derivative. Suppose we want to solve f(x) = 0. Taylor expand f(x) around the current iterate (a)

    f(x) =~ f(a) + f'(a)(x-a)

We want f(x) = 0 so,

    0 = f(a) + f'(a)(x-a)

Multiply both sides by the inverse of f'(x):

    0 = f'(a)^-1 f(a) + x-a

So:

    x = a - f'(a)^-1 f(a)

This is the update equation for Newton's method where a is the current iterate and x is the next iterate.

If f is a multi dimensional function f : R^n -> R^n then the derivative f'(a) is the Jacobian, and inversion becomes matrix inversion.

When we use Newton's method for minimisation of a function g we solve g'(x) = 0, so we pick f(x) = g'(x). Since the formula above contains f' we get a second derivative. The second derivative in multiple dimensions is the hessian.

I just took the index of the data point (i.e. 1, 250) as the y-axis with the intention of "Stretching out" the data set along the y-axis. Otherwise the data would otherwise be illegibly compressed on a 1-D number line.

In retrospect, I could have better represented the data with 2 overlying histograms, but this (somewhat) captures the intent of showing that "more expensive houses tend to have more than 2 bathrooms".

Thank you. I err'ed on the side of caution to (try) to reduce information overload, but I also didn't want to assume too much prior knowledge. Delicate balance...

Maybe a linked post that dives deeper into the 5 steps of Newton's Method? Would that be more approachable?

The comment right below has 2 links:

Gradient Descent explored similarly to this post: http://thelaziestprogrammer.com/sharrington/math-of-machine-...

The difference between the two: https://www.quora.com/In-optimization-why-is-Newtons-method-...

To elaborate a bit: Newton’s method is amazing. But it’s also dangerous and difficult to use. You can talk about it for a long time and use it as the centrepiece of an course on vector analysis. This treatment does not explain anything - it might as well just say “we use the magical newton method, a discussion of which is beyond the scope of this work”. I’d find that honest.

Back when I was in high school Newton's Method was typically introduced as a method to find roots of a polynomial, usually in "Algebra 2" or Trigonometry or about then. That's because there is an easy to grasp formula for it that can be used in the context of a very familiar mathematical construct (polynomials) and produced from differentiation that doesn't technically require an understanding of calc to use (as opposed to actually understand, derive, etc.). I still use it pre-calc when I tutor younger students because it's a great exploratory tool.

I am a programmer trying to learn math, so are my intended audience members.

That said, I should include facts related to convergence, and maybe even speed compared to SGD.

As to the reciprocal -> inverse generalization, do you have any resources you could point we towards to better understand this?

Additionally, a concrete answer to "Why would following the tangent repeatedly be a good idea?" has been hard to come by for me. I am able to visualize this, but if you have resources that explain this well please share.