Why Machines Learn

Why Machines Learn: The Elegant Math Behind Modern AI, Anil Ananthaswamy, 2024

November/December…, 2024

Contents

Context and Reflection

I am reading this book as part of a small book club: The 26-minute Book Club. This sort of book is really not my cup of tea — it is about the math with which various learning algorithms are implemented. I am, at my deepest, interested in the algorithms. Math is neither a strong point nor a deep interest — it seems more like magic to me: You generate a bunch of tautologies, define some terms and rules, and then elaborate it, and all of a sudden it enables you to do things.

I had imagined I would give this a one-meeting trial, and then bow out. However, I found the group (4 other people) quite pleasant, and composed of — barring the one person I know who invited me — interesting strangers that I would not otherwise get to know. And that, especially in retirement, is actually a pretty valuable thing. Everybody else I encounter is mostly a part of my social network.

Another advantage is that working through this book forces me to think about something that I would otherwise ignore: basic math including linear algebra, vectors, a little calculus, probability, gaussian distributions, Bayes’ Theorm. Perhaps it will prove tractable enough that I will feel emboldened to take on fluid dynamics, something that I suspect would help me better understand a whole range of natural patterns, from meteorological to geological…. Or perhaps I will figure out how to distill the ‘qualitative’ aspects that I am interested from the mangle of symbols. We shall see.

C1: Desperately Seeking Patterns

Konrad Lorenz and duckling imprinting. Ducklings can imprint on two moving red objects, and will follow any two moving objects of the same color; or two moving objects of different colors, mutatis mutandis. How is it that an organism with a small brain and only a few instants of exposure can learn something like this?
This leads into a brief introduction to Frank Rosenblatt’s invention of Perceptrons in the late 1950’s. Perceptrons are one of the first ‘brain-inspired’ algorithms that can learn patterns by inspecting labeled data.
Then there is a foray into notation: linear equations (relationships) with weights (aka coefficients) and symbols (variables). Also sigma notation.
McCulloch and Pitt, 1943, story of their meeting, and their model of a neuron — a neurode — that can implement basic logical operations.
The MCP Neurode. Basically, the neurode takes inputs, combines them according to a function, and outputs a 1 if they are over a threshold theta, and a 0 otherwise. If you allow inputs to be negative and lead to inhibition, as well as allow neuroses [sometimes spell-correct is pretty funny] to connect to one another, you can implement all of boolean logic. The problem, however, is that the thresholds theta must be hand-crafted.
Rosenblatt’s Perceptron made a splash because it could learn its weights and theta from the data. An early application was to train perceptrons to recognize hand drawn letters — and it could learn simply by ‘punishing’ it for mistakes.
Hebbian Learning: Neurons that fire together, wire together. Or, learning takes place by the formation of connections between firing neurons, and the loss or severing of connections between neurons that are not in sync.
The difference between the MCP Neurode and Perceptrons is that perceptrons input’s don’t have to be 1 or 0 — they can be continuous. And they are weighted, and they are compared to a bias.
The Perceptron does make one basic assumption: that there is a clear, unambiguous rule to learn — no noise in the data. If this is the case, it can be proven that a perceptron will always find a linear divide (i.e. when there is one to be found).

C2: We are All Just Numbers

Hamilton’s discovery of quaternions, and his inscription on Brougham bridge in Dublin. i² = j² = k² = ijk = -1 Quaternions don’t concern us, but Hamilton developed concepts for manipulating them that are quite important: vectors and scalars.
Scalar/Vector math: computing length; sum; stretching a vector by scalar multiplication;
Dot product: a.b = a1b1 + a2b2 (the sum of the products of the vector’s components). The dot product (AKA the scalar product) is an operation that takes two vectors and returns a single number (a scalar). It’s a way to quantify how much two vectors “align” with each other — that is, the degree to which they point in the same direction.
- E.g. Imagine pushing a model railroad car along some tracks. If you push in the exact direction that the tracks go, all the force you apply goes into moving the car; if you push at an angle to the tracks, only a portion of the force you apply goes into moving the car. This (the proportion of force moving the car along the tracks), is what the dot product gives you.
Something about dot products being similar to weighted sums, which can be used to represent perceptrons??? Didn’t understand this bit. [p. 36-42]
A perceptron is essentially an algorithm for finding a line/plane/hyperplane that accurately divides values into appropriately labeled regions.
Using matrices to represent vectors. Matrix math. Multiplying matrix A with the Transpose of Matrix B
So the point of all this is to take Rosenblatt’s Perceptron and transform it into formal notation that linearly transforms an input to an output.
“Lower bounds tell us about whether something is impossible.” — Manuel Sabin
Minsky and Papert’s book, Perceptrons, poured cold water on the field by proving that Perceptrons could not cope with XOR. XOR could only be solved with multiple layers of Perceptrons, but nobody knew how to train anything but the top layer
I am not clear on why failure to cope with XOR was such cold water…
➔ Later: It is because XOR is a simple logical operation; the inability of Perceptrons handling it suggested that they would not work for even moderately complex problems. Some also generalized the failure to all neural networks, rather than just single layer ones.
Multiple layers requires back-propagation…

C3: The Bottom of the Bowl

McCarthy, Minsky, Shannon and Rochester organized the 1955 Dartmouth summer seminar on Artificial Intelligence. Widrow attended this seminar, but decide it would take at least 25 years to build a thinking machine,
Widrow worked on filtering noise out of signals: He worked on adaptive filtering, meaning a filter that could learn from its errors. Widrow worked on continuous signals; others applied his approach to filtering digital signals. Widrow and Hoff — Adaptive filtering — invented Least Mean Squares algorithm.
Least Mean Squares is a method for quantifying error. What Widrow wanted to do was to create an adaptive filter that would learn in response to errors — this required a method for adjusting parameters of the filter so as to minimize errors. This is referred to as The Method of Steepest Descent, discovered by the French mathematician, Cauchy.
Much of the rest of the chapter introduces math for ‘descending slopes.’ dx/dy moves us along a gradient… the minimum will have a slope of zero. When we have planes or hyperplanes we need to take multiple variables into account so we have partial derivatives.

“If there’s one thing to take away from this discussion, it’s this: For a multi-dimensional or high-dimensional function (meaning, a function of many variables), the gradient is given by a vector. The components of the vector are partial derivatives of that function with respect to each of the variables.
…
What we have just seen is extraordinarily powerful. If we know how to take the partial derivative of a function with respect to each of its variables, no matter how many variables or how complex the function, we can always express the gradient as a row vector or column vector.

Our analysis has also connected the dots between two important concepts: functions on the one hand and vectors on the other. Keep this in mind. These seemingly disparate fields of mathematics-vectors, matrices, linear algebra, calculus, probability and statistics, and optimization theory (we have yet to touch upon the latter two) – will all come together as we make sense of why machines learn.”

…Reading Break…

So with an adaptive filter, you filter the input, and look at the error in the output, and feed that error back into the filter, which adjusts itself to minimize the error.
So first you need to be able to have an input where you already know what the true signal is, so that you can determine the error after the filter has transformed the input. How do you get that? ➔ Later: In the application we’re talking about, this is the training phase. Once the model is trained, you assume the characteristics of the noise will not change and the model will continue to work.
One issue is whether the noise in the signal is always of the same sort — that is, if you train an adaptive filter on a bunch of inputs whose signals you know, will that give you a good chance of having a filter that can appropriately transform an unknown signal? The book uses the example of two modems communicating over a noisy line, and it makes sense (I think) that noise would have fairly uniform characteristics, at least for the session. But that seems unlikely to hold for everything. Can we assume that the noise, in a particular situation, is always the same ,or at least has the same statistical properties?
Suppose the source or nature of the noise in the signal changes over time? Well, you could embed some kind of known signal into the input (I imagine, say, a musical chord), and let the filter learn to adjust the output so that the known chord comes through. But will a filter that preserves the chord also preserve the other information in the signal? I have no idea. I’d think it would depend a lot on (1) the nature of the signal, and (2) the nature of the noise.
I’m confused about the part about adding delays to signal… and I’m confused about how, in real life, you know what the desired signal is.
➔ Later: Still not very clear on the noise issue, but I’m guessing it depends on what you’re applying it to. If the noise is varies in an unpredictable way for a particular application, then the filter/neuron simply won’t work and will produce gibberish.
Anyway, let’s assume we know the desired signal (and hence the noise/error) — how do we quantify the later? We don’t want to just add it up because it can have positive or negative values which would cancel one another out — instead, the errors are squared, and you take their mean to quantify the noise: the is called the Mean Square Error. It is also the case that squaring the errors exaggerates the effects of the larger errors, which seems like a desirable thing.
The math shows that the formula for the error associated with an adaptive filter is quadratic, meaning that it will be concave, and that thus the minimum error will be the minimum of the function. That can be found in multiple ways, either by finding the point at which the slope of the function is zero, or using gradient descent to find it.
A problem is that to do this, you need more and more samples of xn and yn and dn to calculate parameters, and you need to use calculus to calculate partial derivatives, and especially in high dimensional space this becomes burdensome (or impossible).
The solution was that Widrow and Hoff found a (IMO kludgy) way to just estimate the error without doing a lot of work.

weight_New = weight_Old + 2 • <step-size> • <error-for-a-single-point>

This is called the Least Mean Squares (LSM) algorithm.
➔ Later: What they are doing is taking a single data point(a single input-out put pair) at random and using that to estimate the gradient and adjust the weight. Each update a new pair is randomly selected, and over time the algorithm noisily decreases the error. This is called Stochastic Gradient Descent. There is an alternative to this approach called mini-batch gradient descent that uses a randomly-selected set of points (e.g. 32 of them) for each update.

C4: In All Probability

The Monty Hall problem. There are three doors, one of which has a valuable prize behind it, and the others which have only goats. After you’ve picked door 1, Monty opens door 2, revealing a goat. You now have a change to change your pick — should you do that?
The answer. The answer is “yes.” For a long time this seemed counterintuitive to me (and Paul Erdos): revealing what is behind one of the doors should not change the probability of what is behind the other doors. What was tripping me up (ironically) is that I was ignoring the psychology. The key is that Monty is not opening a door at random: he knows what is behind each door, and in particular, he will not open a door that has the prize behind it (as that would destroy the game). So when Monty opens door 2, he is sometimes providing information about both door 2 and about door 3.
Let’s suppose I’ve picked the first door. There are three cases:
(1) Pxx — if I have the correct door, Monty can open either of the others.
(2) xPx — If a goat is behind 1, and P behind 2, Monty can only open 3
(3) xxP — If a goat is behind 1, and P is behind 3, Monty can only open 2
In 2 of these 3 cases, switching to the remaining unopened door gets me the prize. Monty has change the prior probabilities, and so we much re-evaluate.
This argument will hold for any number of doors, because Monty always knows where the car is, and since he will avoid opening that door, every door he opens changes the priors — i.e. gives additional information about the unopened doors.
Later: If we construct a different version of the problem, where, before Monty can pick a new door, an earthquake strikes and door 2 happens to collapse, revealing the goat, there is no reason to change (or not change) your pick. The revelation of the goat behind door 2 does not give us further information about what is behind any of the other doors, since the earthquake’s ‘revelation’ was truely a random event.
Bayes Theorem history. Interestingly, Thomas Bayes’ essay describing his approach was only presented to the Royal Society in 1763, two years after his death, by his friend Richard Price (who later scholars believe made substantive contributions, although Price attributed it all to Bayes).

P (X-is-true | given Evidence-for-X is positive)
IS EQUAL TO
P-X-in-the-world • How-strong-the-evidence-is (e.g. the accuracy of the test)
————————————- (DIVIDED BY) —————————————————-
(P-X-in-the-world • probability of a true positive)
• (1 – P-X-in-the-world) • (1 – probability of a true positive)

The empirical probability in the world * the predictive accuracy given evidence
————————————————————————————————————————–
the likelihood of the world producing that evidence
(=sum of probabilities of all ways of producing that evidence)

…Reading Break…

Machine Learning is inherently probabilistic because there are an infinite number of hyperplanes that can discriminate between a learned alternative, and it has settled on one of them for no particular reason. Other factors that make ML less than accurate are that the data itself may have errors, and that the amount of data drawn upon is limited. Later: And we must keep in mind that ML is only minimizing error — whether the result has enough signal to be useful is an empirical and domain-dependent issue.
Distinction between theoretical probability and empirical probability (e.g., theoretical probability of a fair coin coming up heads is 50%; empirical probability of a fair coin coming up heads depends on actually doing it, and it will approach but not reach the theoretical probability as one increases the number of empirical samples.
Aside: There is also the issue of the degree to which real-world events are actually expressions of mathematical distributions. It seems elegant to assume that, but is it really so?
The case of a coin flip is an example of a Bernoulli distribution. It has only two values, a and b, and can be characterized by a probability p that such that p is the probability of a, and (1–p) is the probability of b.
Distributions with a mean and a variance (aka standard deviation). Now consider the case where you have N>2 outcomes, each with their own probability. This is distribution can be characterized by a mean and standard deviation — the mean (aka the expected value) is the sum of the values of each outcome multiplied by their probabilities, and the standard deviation is the square root of {the sum of the squares of the (deviations of each value from the mean)} — or the sum of the absolute value of each values difference from the mean. We can talk about the distribution as a whole in terms of its probability mass function.
So far we’ve been talking about variables with discrete values, but we can instead talk about variables with continuous values. Here we can’t talk about the probability of a particular value because there are an infinite number of values and the probability of any single perfectly precise value is zero. So, instead, what we do is talk about the probability of a value occurring within particular bounds: this is called the probability density function. [Aside: But wouldn’t it be possible to do some calculus like move where we look at what happens as a finite interval approaches zero?]
The important point is that whether we have a variable with discrete or continuous values, we can use well-known and analytically well understood functions to characterize the distribution.
Machine Learning. Let’s being with a set of labeled data points: y is a label that has two values, and x is an instance of the data. y is categorical; x is a vector with N components. This data set can be represented as a matrix: y₁, x₁₁, x₁₂, x^₁₃ .. x_1n and so on for y₂, y₃, etc. Each component of x is a feature that the algorithm will use to discriminate which y x belongs to.
Now, if you knew the underlying distribution P(y, x), you could determine the probability of y=A given x, and the probability of y≠A given x, and use the highest probability to assign the label. If this were the case, this is what would be termed a Bayes Optimal Classifier. [Aside: I’m a little unclear on this — it seems like it’s dependent on a particular situation, and so it seems odd to give it this sort of name.]
But usually you don’t know the underlying distribution, and so it must be estimated. Often it is easier to make assumptions about the underlying distribution (Bernoulli? Gaussian?), but it is important to keep in mind that these are idealizations chosen to make the math easier.
Aside: A Gaussian distribution is defined as being symmetric with respect to a single mode (which is also the mean and median), with asymptotic tails that never reach 0.
There are two approaches to estimating distributions:
One is MLE or Maximum Likelihood Estimation, involves selecting a theta (that is a distribution with particular parameters indicated by theta) that maximizes the likelihood of observing (generating?) that labeled data you already have.
In the text, MLE is exemplified by imagining a set of data about two populations’ heights, labeled short or tall, and that each population has a gaussian distribution, and thus that all the data will be best modeled by a combination of the two distributions. ???But is that still a gaussian distribution? And what is the rational for choosing a gaussian distribution rather than some other distribution???
The other is called MAP, for Maximum A Posteriori estimation. As best I can tell, this involves estimating the distribution based on our experience of the world and representing it as a function. Then you set the derivative to zero, and solve that equation (after checking to be sure that you’ve got the maximum rather than the minimum). If this approach does not yield a ‘closed’ solution (most of the time), you can take the negative of the function and use gradient descent to find its minimum.
MLE vs MAP. So MLE tries to find the maximum using the data, and MAP tries to find the maximum of a distribution you’ve guessed at. MLE works best when there’s a lot of data; MAP works best when there isn’t much data you can make a reasonable guess about the underlying distribution).
The Federalist Papers example: which unattributed papers were written by Madison and which by Hamilton. A first approach was to analyze the length of sentences in papers know to be written by each author, and use the mean length and the standard deviation to discriminate: unfortunately the means and SDs for each author were almost identical. Later, someone suggested using the frequency of word use, and, in particular, function words tended to reliably discriminate between the known works of the two authors: this was then used to predict which of the unattributed essays were written by which author.
Aside: It is not clear to me whether the achievement of Mosteller and Wallace was due to the use of Bayesian reasoning, or to the realization the word choice was a very good discriminator between the authors. ➔ LATER: Consensus among scholar from many fields indicates, according to chatGPT, that their work did indeed validate the use of Bayesian inference, as well as creating a new approach to linguistics, and indicating the ways in which computers could be used.
The Penguin example.
A trick that statisticians use is to assume that the distributions for each feature under consideration are independent — seems like a bit of a leap, but it appears to work and it makes the math easier and requires less data.
Naive or Idiot Bayesian classifier.

… reading break…