About Me
moon indicating dark mode
sun indicating light mode

Demystifying Deep Learning: Part 2

Linear and Logistic Regression

July 29, 2018

3 min read


Today, we get our hands dirty with our first machine learning algorithms! Whilst we are starting out simple, these algorithms are still very useful and can be used as a baseline to measure the later machine learning algorithms we cover - to ensure the algorithm is behaving as we expect.


Often when we have a set of data points in science, we want to draw a straight line of best fit through the points (see picture above). Linear Regression involves learning this straight line of best fit, only with more dimensions - instead of axes xx and yy, we have axes x1,x2,...,xnx_1, x_2, ... ,x_n and yy.

In Linear Regression, we use this straight line of best fit and the values for x1,x2,...,xnx_1, x_2, ... ,x_n, which we call input features to predict the value of the output yy. This is useful in regression tasks, e.g. housing prices.

Now instead of a real number output, suppose we wanted to classify this input as one of two classes e.g. fraudulent vs non-fraudulent credit card data. Logistic Regression involves taking this output and applying a decision function to this - we end up with a probability - if close to 1 the input is one class, and if closer to 0 the input is the other class. We can think of this probability as the “confidence” the algorithm has.

Side note: The name logistic regression is a little confusing as it is a classification algorithm, however this is because the term regression in the context of machine learning differs from regression in statistical maths. For our purposes, we’ll use the definition in the first blog post of the series - of real-valued output.


The equation for linear regression extends naturally from the equation of a straight line:

y=mx+cy= mx + c

Just like how mm is the gradient of the line, we assign weights to the input features - let wiw_i denote the weight corresponding to feature xix_i - you can loosely think of wiw_i as the gradient of the line in the direction of the xix_i axis.

Another way of thinking of fitting a line of best fit is as projecting the original data points yy onto the line, denoting our projection as y^\hat{y}. Extending this to nn dimensions, instead of a line we consider a n-dimensional hyperplane that we are projecting our data points onto.

So the equation for linear regression is:

y^=j=1nwjxj+b\hat{y} = \sum_{j=1}^{n}{w_jx_j} + b

For logistic regression, we apply the sigmoid function σ(x)\sigma(x) to this output - this scales the values to between 0 and 1 (to get the probability) - see the graph below:

Sigmoid graph

σ(x)=11+ex\sigma(x) = \frac{1}{1+e^{-x}}

y^=σ(j=1nwjxj+b)\hat{y} = \sigma(\sum_{j=1}^{n}{w_jx_j} + b)


One convention we will follow is (x(i),y(i))(x^{(i)}, y^{(i)}) refers to the ithi^{th} training example, with y^(i)\hat{y}^{(i)} the corresponding prediction by the model. In general (i)^{(i)} refers to the value for ithi^{th} training example. We use mm to denote the number of examples in the training set and nn to denote the number of input features.


We will also represent our operations as matrix multiplications, since it is cleaner to express and also more efficient when implemented compared to a for loop across the values.

So here, we store our input in a n x m matrix XX and the labels as a 1 x m matrix YY. Then we can represent WW as a 1 x n matrix and express the calculation as:

Y^=W.X+b\hat{Y} = W.X + b - note Y^\hat{Y} has same dimensions as labels YY.

In general, if we have an equation of form Cij=kAikBkjC_{ij} = \sum_k A_{ik} B_{kj} this corresponds to a matrix multiplication C=A.BC = A.B

(Here we have Y^1i=j=1nW1jXji\hat{Y}_{1i} = \sum_{j=1}^{n}{W_{1j}X_{ji}})

Join me on this learning journey!

This summer I’m using my blog to teach the topics I’ve learnt this year. It’s a win-win - you get computer science tutorials and I get to share it with you!


We will be writing the code snippets in Python and we are using Numpy (a linear algebra library) to carry out the maths operations.

Our motivating problem for linear regression is the Boston Housing Prices dataset, and for logistic regression the Breast Cancer dataset (classifying tumour as benign/malignant). See the Jupyter notebook to see the input preprocessing.

The intention of the blog posts is to highlight the pertinent parts of the code - the notebook accompanying the blog post can be used to view the rest of the code and implement the algorithm on a concrete problem to consolidate.

With all that said, let’s dive into the code - note how np.dot is used for matrix multiplication.

import numpy as np #we use numpy to do the maths operations
#_lin refers to linear regression, _log for logistic regression
#similarly _train for training set, _test for test set.
#initialise weights and bias to random values
# we'll get to training them in the next blog post.
W_lin = np.random.randn(Y_lin_train.shape[0], X_lin_train.shape[0]) #1xm matrix
b_lin = np.random.randn(Y_lin_train.shape[0],1) #1x1 matrix
W_log = np.random.randn(Y_log_train.shape[0], X_log_train.shape[0])#1xm matrix
b_log = np.random.randn(Y_log_train.shape[0],1)) #1x1 matrix
#Linear Regression
def forward_lin(X, W, b):
return np.dot(W,X)+b
def sigmoid(z):
return 1.0/(1+np.exp(-z))
#Logistic regression
def forward_log(X,W,b):
return sigmoid(forward_lin(X,W,b))


At this point, you’re probably thinking, “Great! But this is just maths, where is the learning?“. Hold tight, because in the next blog post we’ll be doing just that! We will look at how exactly most machine learning algorithms even “learn”, and how we evaluate their performance.

Share This On Twitter

If you liked this post, please consider sharing it with your network. If you have any questions, tweet away and I’ll answer :) I also tweet when new posts drop!

PS: I also share helpful tips and links as I'm learning - so you get them well before they make their way into a post!

© Mukul Rathi 2020