2022-04-20-gbo-ml-basics-1.md

layout	title	tags	mathjax
post	GBO notes: Machine learning basics (Part 1)	gbo ml-basics	true

In this series of notes we will review some basic concepts that are usually covered in an Intro to ML course. These are based on this course from Cornell.

In Part 1, we will look at the basic problem of supervised learning, simple classifiers such as k-NN and perceptron, and MLE and MAP estimators.

Machine learning setup

In supervised learning, we are given a dataset $D = {(\mathbf{x}_1,y_1),\ldots,(\mathbf{x}_n,y_n)} \in \mathbb{R}^d\times \mathcal{C}$, and the goal is to learn a function $h$ such that $h(\mathbf{x}) = y$ with high probability. Here $\mathbf{x}_i,y_i)$ are sampled from some unknown distribution.

We usually have to make some assumption about the type of function $h$ so that learning is possible. This is called the no free lunch theorem.

Once we have selected the hypothesis class, we find the best function $h$ within the class by solving some kind of optimization problem. The objective of this optimization is to minimize some notion of loss on the training data.

To ensure that we are generalizing and not just memorizing the training data, we keep a held-out validation set on which we periodically measure the loss while training. If the test data is large enough, it is an unbiased estimator of the model performance, because of the weak law of large numbers.

k-nearest neighbors

It is a non-parametric classification method. The idea is that given a test input $\mathbf{x}$, return the most common label among its $k$ nearest neighbors ($k$ most similar training inputs). Small values of $k$ will return noisy result, and very large values will increase computation. Usually, the $L_p$-norm is used as the distance metric to compute the $k$-nearest neighbors.

Theorem: For large $n$, the 1-NN error is no more than twice the error of a Bayes optimal classifier.

The proof uses the fact that for large $n$, the nearest neighbor is the same point as the test point. Unfortunately, in high-dimensional setting, this is never true, since two points in high-dimensional space are far apart.

However, if we consider a hyperplane in the high-dimensional space, the distance of points to the hyperplane does not change regardless of the dimensionality! This means that in high dimensions, distances between points are large but distances to hyperplanes are tiny. Classifiers such as SVMs and perceptrons use this idea.

Perceptron

It is used for binary classification and assumes that the data is linearly separable, i.e., separable by a hyperplane. The hypothesis class is defined as $h(\mathbf{x}_i) = \textrm{sign}(\mathbf{w}^T \mathbf{x}_i + b)$. We can absorb the bias into the weight by adding a constant dimension to the input.

To learn $\mathbf{w}$, we keep updating it as $\mathbf{w} = \mathbf{w} + y\mathbf{x}$ until it learns to classify everything correctly.

Theorem: Suppose there is a hyperplane separating the points in a unit-normalized hypersphere, defined by the normal $\mathbf{w}*$. Let $\gamma$ be its margin, or the distance to the nearest point. Then the perceptron algorithm makes at most $\frac{1}{\gamma^2}$ mistakes.

Estimating probabilities from data

Maximum likelihood estimate (MLE): We first assume some distribution from which the data was sampled from, and then compute the parameters of the distribution which maximizes the probability of generating the data, i.e.,

$$ \theta^{MLE} = \textrm{arg}\max_{\theta} P(X;\theta). $$

Example: for estimating the probability of head in a coin toss given some observations, we can assume that the probability comes from a Bernoulli distribution with parameter $\theta$, and then applying MLE will give us $\theta = \frac{\mathrm{number of heads}}{\mathrm{number of tosses}}$.

However, if we have small number of observations, this method of estimating $\theta$ can give a very biased estimate. Instead, we can do it the Bayesian way. Assume that $\theta$ comes from some prior distribution $P(\theta)$. To make computation easier, we can choose a conjugate prior (for example, beta prior for bernoulli distribution and dirichlet prior for multivariate distribution).

Maximum a Posteriori (MAP): The MAP estimate is given as:

$$ \theta^{MAP} = \mathrm{arg}\max_{\theta} P(\theta|D) = \mathrm{arg}\max_{\theta} P(D|\theta)P(\theta). $$

Quick summary:

• MLE prediction: $P(y|\mathbf{x};\theta)$ Learning: $\theta = \textrm{arg}\max_{\theta}P(D;\theta)$. Here $\theta$ is purely a model parameter.
• MAP prediction: $P(y|\mathbf{x},\theta)$ Learning: $\theta = \textrm{arg}\max_{\theta}P(\theta|D) = \mathrm{arg}\max_{\theta} P(D|\theta)P(\theta)$. Here $\theta$ is a random variable.