How to Explain Machine Learning to a Software Engineer
Source: Sebastian Raschka

Software engineering is about developing programs or tools to automate tasks. Instead of "doing things manually," we write programs; a program is basically just a machine-readable set of instructions that can be executed by a computer. Let's consider a classic example: e-mail spam filtering. Assuming that we have access to the source code of our e-mail client and know how to handle it, we could come up with an instinctive set of rules that may help us with our spam problem.


Some code

For example: if not "sender in contacts": if "subject line contains BUY!: e-mail spam folder:" else if ...

It is intuitive to say that coming up with these rules is a pretty tedious task. Needless to say that we have to test our spam filter on real-world data, evaluate and improve it constantly, change and update rules, and so forth. Again, our goal is automation: we want to write a set of instructions that automatically filters out spam e-mails so that we don't have to "manually" delete them from our e-mail inbox.

Now, machine learning is all about automating automation! Instead of coming up with the rules to automate a task such as e-mail spam filtering ourselves, we feed data to a machine learning algorithm, which figures out these rules all by itself. In this context, "data" shall be representative sample of the problem we want to solve -- for example, a set of spam and non-spam e-mails so that the machine learning algorithm can "learn from experience."

Traditional vs machine learning programming paradigms

In "conventional" programming, we code up a set of rules, feed it to the computer together with the data, and hope that it produces the desired results.


traditional programming:
set of rules + data -> computer -> results

In machine learning, we take data (e.g., e-mails), provide information about the desired results (spam and non-spam labels for these e-mails), and feed it to a learning algorithm, which in turn executed by a computer. The computer then learns a set of rules that we can use to automate (solve) our problem task.

machine learning:
results + data -> machine learning algorithm + computer -> set of rules

Or in other words, machine learning is about finding the optimal instructions to automate a task. Machine learning algorithms are instructions for computers to learn other instructions automatically from data or experience. Therefore, machine learning is the automation of automation.