CSC 411: Lecture 02: Linear Regression

Richard Zemel, Raquel Urtasun and Sanja Fidler

University of Toronto

(Most plots in this lecture are from Bishop’s book)

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 1 / 22

Problems for Today
What should I watch this Friday?
Goal: Predict movie rating automatically!
Goal: How many followers will I get?
Goal: Predict the price of the house

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 2 / 22

Regression
What do all these problems have in common?
I Continuous outputs, we’ll call these t
(e.g., a rating: a real number between 0-10, # of followers, house
price)
Predicting continuous outputs is called regression
What do I need in order to predict these outputs?
I Features (inputs), we’ll call these x (or x if vectors)
I Training examples, many x (i) for which t (i) is known (e.g., many
movies for which we know the rating)
I A model, a function that represents the relationship between x and t
I A loss or a cost or an objective function, which tells us how well our
model approximates the training examples
I Optimization, a way of finding the parameters of our model that
minimizes the loss function
Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 3 / 22
Today: Linear Regression

Linear regression
I continuous outputs
I simple model (linear)

Introduce key concepts:

I loss functions
I generalization
I optimization
I model complexity
I regularization

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 4 / 22

Simple 1-D regression

Circles are data points (i.e., training examples) that are given to us
The data points are uniform in x, but may be displaced in y
t(x) = f (x) + 
with  some noise
In green is the ”true” curve that we don’t know
Goal: We want to fit a curve to these points
Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 5 / 22
Simple 1-D regression

Key Questions:
I How do we parametrize the model?
I What loss (objective) function should we use to judge the fit?
I How do we optimize fit to unseen test data (generalization)?

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 6 / 22

Example: Boston Housing data
Estimate median house price in a neighborhood based on neighborhood
statistics
Look at first possible attribute (feature): per capita crime rate

Use this to predict house prices in other neighborhoods

Is this a good input (attribute) to predict house prices?
Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 7 / 22
Represent the Data
Data is described as pairs D = {(x (1) , t (1) ), · · · , (x (N) , t (N) )}
I x ∈ R is the input feature (per capita crime rate)
I t ∈ R is the target output (median house price)
(i)
I simply indicates the training examples (we have N in this case)
Here t is continuous, so this is a regression problem
Model outputs y , an estimate of t

y (x) = w0 + w1 x

What type of model did we choose?

Divide the dataset into training and testing examples
I Use the training examples to construct hypothesis, or function
approximator, that maps x to predicted y
I Evaluate hypothesis on test set
Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 8 / 22
Noise

A simple model typically does not exactly fit the data

I lack of fit can be considered noise

Sources of noise:
I Imprecision in data attributes (input noise, e.g., noise in per-capita
crime)
I Errors in data targets (mis-labeling, e.g., noise in house prices)
I Additional attributes not taken into account by data attributes, affect
target values (latent variables). In the example, what else could affect
house prices?
I Model may be too simple to account for data targets

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 9 / 22

Least-Squares Regression

Define a model
y (x) = function(x, w)

Linear: y (x) = w0 + w1 x
Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 10 / 22
Optimizing the Objective
One straightforward method: gradient descent
I initialize w (e.g., randomly)
I repeatedly update w based on the gradient
∂`
w ←w−λ
∂w

λ is the learning rate

For a single training case, this gives the LMS update rule (Least Mean
Squares):
w ← w + 2λ(t (n) − y (x (n) ))x (n)

w ← w + 2λ (t (n) − y (x (n) )) x (n)

| {z }
error

Note: As error approaches zero, so does the update (w stops changing)

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 11 / 22

Optimizing Across Training Set
Two ways to generalize this for all examples in training set:
1. Batch updates: sum or average updates across every example n, then
change the parameter values
N
X
w ← w + 2λ (t (n) − y (x (n) ))x (n)
n=1

2. Stochastic/online updates: update the parameters for each training

case in turn, according to its own gradients

Algorithm 1 Stochastic gradient descent

1: Randomly shuffle examples in the training set
2: for i = 1 to N do
3: Update:

w ← w + 2λ(t (i) − y (x (i) ))x (i) (update for a linear model)

4: end for

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 12 / 22

Analytical Solution?

For some objectives we can also find the optimal solution analytically
This is the case for linear least-squares regression
How?
Compute the derivatives of the objective wrt w and equate with 0
Define:

t = [t (1) , t (2) , . . . , t (N) ]T

1, x (1)
 
 1, x (2) 
X=  ... 


1, x (N)

Then:
w = (XT X)−1 XT t
(work it out!)
Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 13 / 22
Multi-dimensional Inputs

One method of extending the model is to consider other input dimensions

y (x) = w0 + w1 x1 + w2 x2

In the Boston housing example, we can look at the number of rooms

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 14 / 22

Linear Regression with Multi-dimensional Inputs

Imagine now we want to predict the median house price from these
multi-dimensional observations
Each house is a data point n, with observations indexed by j:
 
(n) (n) (n)
x(n) = x1 , · · · , xj , · · · , xd

We can incorporate the bias w0 into w, by using x0 = 1, then

d
X
y (x) = w0 + w j xj = w T x
j=1

We can then solve for w = (w0 , w1 , · · · , wd ). How?

We can use gradient descent to solve for each coefficient, or compute w
analytically (how does the solution change?)

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 15 / 22

More Powerful Models?Fitting a Polynomial

What if our linear model is not good? How can we create a more
complicated model?
We can create a more complicated model by defining input variables that are
combinations of components of x
Example: an M-th order polynomial function of one dimensional feature x:
M
X
y (x, w) = w0 + wj x j
j=1

where x j is the j-th power of x

We can use the same approach to optimize for the weights w
How do we do that?

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 16 / 22

Which Fit is Best?
from Bishop

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 17 / 22

Generalization
Generalization = model’s ability to predict the held out data
What is happening?
Our model with M = 9 overfits the data (it models also noise)
Not a problem if we have lots of training examples
Let’s look at the estimated weights for various M in the case of fewer
examples
The weights are becoming huge to compensate for the noise
One way of dealing with this is to encourage the weights to be small (this
way no input dimension will have too much influence on prediction). This is
called regularization

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 18 / 22

Regularized Least Squares
Increasing the input features this way can complicate the model considerably
Goal: select the appropriate model complexity automatically
Standard approach: regularization
N
X
˜
`(w) = [t (n) − (w0 + w1 x (n) )]2 + αwT w
n=1

Intuition: Since we are minimizing the loss, the second term will encourage
smaller values in w
When we use the penalty on the squared weights we have ridge regression in
statistics
Leads to a modified update rule for gradient descent:
N
X
w ← w + 2λ[ (t (n) − y (x (n) ))x (n) − αw]
n=1

Also has an analytical solution: w = (XT X + α I)−1 XT t (verify!)

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 19 / 22
Regularized least squares

Better generalization
Choose α carefully

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 20 / 22

1-D regression illustrates key concepts

Data fits – is linear model best (model selection)?

I Simple models may not capture all the important variations (signal) in
the data: underfit
I More complex models may overfit the training data (fit not only the
signal but also the noise in the data), especially if not enough data to
constrain model
One method of assessing fit: test generalization = model’s ability to predict
the held out data
Optimization is essential: stochastic and batch iterative approaches; analytic
when available

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 21 / 22

So...

Which movie will you watch?

Zemel, Urtasun, Fidler (UofT) CSC 411: 02-Regression 22 / 22