CSC411 Lecture 2
Recall that regression means fitting a model to data. So in 1D regression, you are matching a curve of values to a bunch of data points.
The key questions to keep in mind are: -How do we parammetrize that model? -What loss (objective) function should we use to judge our model? i.e. judge how far off our model is from reality. -How do we generalize our model and its fit to unseen data?
We choose to describe the data D as a bunch of pairs ((x1, t1), (x2, t2), ...) where x is the input feature and t is the target output. Now we can create new data for given values of x by making a pair (x, f(x)), where f(x) is our model.
We can use the training examples we're given to construct a hypothesis, or function approximator, that maps x to predicted y. Then we evaluate this hypothesis on the test set to see how well it works out. If we leave some data out of our training examples, we can use the excess data to figure out how much loss our function has.
There are a few problems that this runs into. We call lack of fit that results from these problems noise:
For least-squares regression, the standard loss function measures the squared error in y from x, in other words the squared difference between our prediction and the objective value.
N = amount of data w = vector of parameters for our model
J(w) = sum(n=1 to N)[tn - (w0 + (w1)(xn))]^2
One straightforward method of minimizing the error, or optimizing the objective, is to initialize w randomly and then repeatedly minimize it based on gradient descent in J
w <- (update to) w - lambda * derivative of J w.r.t. w
where lambda is the learning rate. The lower the learning rate, the longer it takes to get to the minimum, leading to smaller changes in w. The choice of our optimization method and learning rate is very important, and will be examined in more detail later.
The LMS update rule suggests that for a single training case we adjust using the error applied to the derivative.
We have two ways of generalizing this:
Stochastic / online update: Update params for each training case in turn, according to that cases own gradients
Batch update: sum or average updates across every example i, then change the parameter values afterwards.
There's an underlying assumption that the sample is independent and identically distributed(i.i.d.), i.e. the data has no bias. This assumption is applied to make the math for all these algorithms work.
Instead of examining values case by case, we take the derivatives, set to zero, and then solve for the parameters.
SOLVE: dJ(w) / dw0 = 0
Useful to work through this and get the expected result.
An alternative to the least-squares objective is to try and minimize the absolute error (sum over n)|tn - (wT)(xn)|. (What is wT)? Taking the derivative of this isn't easy, so instead you have to solve a linear programming problem, which is more time-consuming.
As we will see later though, there are reasons to want to use this optimization method though.
When we generalize to more than one input, each training case's x value becomes a vector of x1 to xd, representing each of the input values.
Essentially we have a polynomial function with x0 = 1, so in the form
y = w0 + w1x1 + w2x2 + ...
We can try to fit a polynomial function to our regression by defining input vars that are combos of components of x.
Regularization, otherwise known as weight decay, aims to find the simplest model that fits the data the best. We can modify the least-squares method to fit this.