# How Do Machines Learn?

You’ve probably heard of machine learning models that can read human handwriting or understand speech. You might know that these models had to be trained in order to accomplish these tasks– they had to learn. But how exactly does a machine “learn”? What are the steps involved?

In this article, I’m going to be giving a high-level overview of how the “learning” in machine learning happens. I’m going to talk about fundamental ML concepts including cost functions, optimization, and linear regression. I’ll outline the basic framework used in most machine learning techniques.

Data is the foundational of any machine learning model. In a nutshell, the data scientist feeds a bunch of data into the ML model and, as it starts to “learn” from the data, the model will eventually develop a solution. What is the solution? The solution is typically a function that describes the relationship in the data. For a given input, the function should be able to provide the expected output.

In the case of linear regression, one of the most basic ML models, the regression model “learns” two parameters: the slope and the intercept. Once the model learns these parameters to the desired extent, the model can be used to compute the output y for a given input X (in the linear regression equation y = b0 + b1*X). If you’re unfamiliar with linear regression, take a look at my article on linear regression to understand this better.

So now that we know what the goal of machine learning is, we can talk about how exactly the learning happens. The machine learning model usually follows three core steps in order to “learn” the relationship in the data as described by the solution function:

1. Predict
2. Calculate the error
3. Learn

The first step is for the model to make a prediction. To start, the model may make arbitrary guesses for the values that it is solving for in the solution function. In the case of linear regression, the ML model would make guesses for the values of the slope and intercept.

Next, the model would check its prediction against the actual test data and see how good/bad the prediction was. In other words, the model calculates the error in its prediction. In order to compare the prediction against the data, we need to find a way to measure how “good” our prediction was.

Finally, the model will “learn” from its error by adjusting its prediction to have a smaller error.

The model will repeat these 3 steps– predict, calculate error, and learn– a bunch of times and slowly come to the best coefficients for the solution. This simple 3-step algorithm is the basis for training most machine learning models.

When I talked about calculating error earlier, I didn’t talk about the ways in which we measure how “good” or “bad” our predictions are. That leads me to the next topic: cost functions. In machine learning, a cost function is a mechanism that returns the error between predicted outcomes and the actual outcomes. Cost functions measure the size of the error to help achieve the overall goal of optimizing for a solution with the lowest cost.

The objective of an ML model is to find the values of the parameters that minimize the cost function. Cost functions will be different depending on the use case but they all have this same goal.

The Residual Sum of Squares is an example of a cost function. In linear regression, the Residual Sum of Squares is used to calculate and measure the error in predicted coefficient values. It does this by finding the sum of the gaps between the predicted values on the linear regression line and the actual data point values (check out this article for more detail). The lowest sum indicates the most accurate solution.

Cost functions fall under the broader category of optimization. Optimization is a term used in a variety of fields, but in machine learning it is defined as the process of progressing towards the defined goal, or solution, of an ML model. This includes minimizing “bad things” or “costs”, as is done in cost functions, but it also includes maximizing “good things” in other types of functions.

In summary, machine learning is typically done with a fundamental 3-step process: make a prediction, calculate the error, and learn / make adjustments. The error in a prediction is calculated using a cost function. Once the error is minimized, the model is done “learning” and is left with a function that should provide the expected result for future data.

# Introduction to Linear Regression

In this article, I will define what linear regression is in machine learning, delve into linear regression theory, and go through a real-world example of using linear regression in Python.

### What is Linear Regression?

Linear regression is a machine learning algorithm used to measure the relationship between two variables. The algorithm attempts to model the relationship between the two variables by fitting a linear equation to the data.

In machine learning, these two variables are called the feature and the target. The feature, or independent variable, is the variable that the data scientist uses to make predictions. The target, or dependent variable, is the variable that the data scientist is trying to predict.

Before attempting to fit a linear regression to a set of data, you should first assess if the data appears to have a relationship. You can visually estimate the relationship between the feature and the target by plotting them on a scatterplot.

If you plot the data and suspect that there is a relationship between the variables, you can verify the nature of the association using linear regression.

### Linear Regression Theory

Linear regression will try to represent the relationship between the feature and target as a straight line.

Do you remember the equation for a straight line that you learned in grade school?

y = mx + b, where m is the slope (the number describing the steepness of the line) and b is the y-intercept (the point at which the line crosses the vertical axis)

Equations describing linear regression models follow this same format.

The slope m tells you how strong the relationship between x and y is. The slope tells us how much y will go up or down for a given increase or decrease in x, or, in this case, how much the target will change for a given change in the feature.

In theory, a slope of 0 would mean there is no relationship at all between the data. The weaker the relationship is, the closer the slope is to 0. But if there is a strong relationship, the slope will be a larger positive or negative number. The stronger the relationship is, the steeper the slope is.

Unlike in pure mathematics, in machine learning, the relationship denoted by the linear equation is an approximation. That’s why we refer to the slope and the intercept as parameters and we must estimate these parameters for our linear regression. We even use a different notation in which the intercept constant is written first and the variables are greek symbols:

Even though the notation is different, it’s the exact same equation of a line y=mx+b. It is important to know this notation though because it may come up in other linear regression material.

But how do we know where to make the linear regression line when the points are not straight in a row? There are a whole bunch of lines that can be drawn through scattered data points. How do we know which one is the “best” line?

There will usually be a gap between the actual value and the line. In other words, there is a difference between the actual data point and the point on the line (fitted value/predicted value). These gaps are called residuals. The residuals can tell us something about how “good” of an estimate our line is making.

Look at the size of the residuals and choose the line with the smallest residuals. Now, we have a clear method for the hazy goal of representing the relationship as a straight line. The objective of the linear regression algorithm is to calculate the line that minimizes these residuals.

For each possible line (slope and intercept pair) for a set of data:

1. Calculate the residuals
2. Square them to prevent negatives
3. Add the sum of the squared residuals

Then, choose the slope and intercept pair that minimizes the sum of the squared residuals, also known as Residual Sum of Squares.

Linear regression models can also be used to estimate the value of the dependent variable for a given independent variable value. Using the classic linear equation, you would simply substitute the value you want to test for x in y = mx + b; y would be the model’s prediction for the target for your given feature value x.

### Linear Regression in Python

Now that we’ve discussed the theory around Linear Regression, let’s take a look at an example.

Let’s say we are running an ice cream shop. We have collected some data for daily ice cream sales and the temperature on those days. The data is stored in a file called temp_revenue_data.csv. We want to see how strong the correlation between the temperature and our ice cream sales is.

import pandas
from pandas import DataFrame

X = DataFrame(data, columns=['daily_temperature'])
y = DataFrame(data, columns=['ice_cream_sales'])

First, import Linear Regression from the scikitlearn module (a machine learning module in Python). This will allow us to run linear regression models in just a few lines of code.

from sklearn.linear_model import LinearRegression

Next, create a LinearRegression() object and store it in a variable.

regression = LinearRegression()

Now that we’ve created our object we can tell it to do something:

The fit method runs the actual regression. It takes in two parameters, both of type DataFrame. The feature data is the first parameter and the target data is the second. We are using the X and y DataFrames defined above.

regression.fit(X, y)

The slope and intercept that were calculated by the regression are available in the following properties of the regression object: coef_ and intercept_. The trailing underline is necessary.

# Slope Coefficient
regression.coef_

# Intercept
regression.intercept_

How can we quantify how “good” our model is? We need some kind of measure or statistic. One measure that we can use is called R2, also known as the goodness of fit.

regression.score(X, y)
output: 0.5496...

The above output number (in percentage) is the amount of variation in ice cream sales that is explained by the daily temperature.

Note: The model is very simplistic and should be taken with a grain of salt. It especially does not do well on the extremes.