ap_spec_machine_learning.tex

\chapter{Special Topic: Machine Learning}

Machine learning\index{machine learning} (ML) 
is a type of artificial intelligence, and as far
as I can tell, what distinguishes it is the use of statistical
methods to try to infer ``hidden" information. Often this hidden
information can be thought of as a set of features common to the
to-be-studied data that are not known by the user at runtime. 
The goal of the ML algorithm, then, is to try to identify these
features.

For example, a common task for ML algorithms is to identify pictures,
say to determine whether a photo contains a traffic light. The algorithm
must identify some features in these photos; maybe it ``looks out" for
photos containing a red, yellow, and green circle, but remember,
you, the user, to do not tell the algorithm what to look for ahead of time.
Mathematically, these features can be thought of as generalized 
models\footnote{In the Bayesian statistical sense.} which can be
represented mathematically, and the goal of the algorithm is to
find the ``best" model parameters.

Broadly, ML algorithms can be divided into {\it supervised
learning}\index{supervised learning}, {\it unsupervised
learning}\index{unsupervised learning}, and {\it reinforcement
learning}\index{reinforcement learning}.

For supervised learning: The model is {\it trained}\index{training} 
on a labeled dataset. 
Heuristically, we think of training as ``teaching" a model based on the
data. In more detail, this is carried out through several steps:
\begin{itemize}
\item Before training, data is properly cleaned and prepared. This could 
involve rectifying missing values, handling outliers, etc. You label
these data, i.e. they are classified ahead of time\footnote{These photos
contain traffic lights; those photos do not.}. We split our original
data into a {\it training set}, which will be used in step 2, and
a {\it validation set}, which will be used in step 4.
\item Next, the data are given as input to the model, which then calculates
a prediction based on these data. It compares that result against the label
and adjusts its parameters accordingly. This is the {\it
learning}\index{learning} part of the algorithm. This comparison is usually
handled by minimizing a {\it loss function}\index{loss function}, which
is the difference between the prediction and actual label.
\item We repeat step 2, each time feeding different subsets of the data, 
slowly adjusting the model. Each such iteration is sometimes called
an {\it epoch}\index{epoch}. This is done many times until the model's 
predictions are satisfactory or we hit some maximum number of epochs. 
\item Now that we have trained our model using the training set, we will
try the model on the validation set. We can measure how good our model
is by, for example, the fraction of the validation set the model calls
correctly. This process is {\it validation}\index{validation}.
\item After the model has been trained and validated, it can be tested on a separate testing dataset to further evaluate its performance.
\end{itemize}

% uhhh actually hard to tell difference between testing and validation... 

%2. Unsupervised Learning: In this case, there's no teacher and the model must figure out patterns from the input data on its own. The analogy for this is like learning to play a new game without the rules and learning the rules as you play the game.
%
%3. Reinforcement Learning: Here, the machine learns by interacting with its environment. The model will learn to perform an action from states within the environment that maximize the reward through a trial and error approach.
%
\bibliographystyle{unsrtnat}
\bibliography{bibliography}