Machine Learning vs non-Machine Learning algorithm

Recently, I was interacting with some non-technical folks and noticed that people use the term ML even when an algorithm doesn’t actually use any Machine Learning.

So here is the simplest definition (that I know of) which explains the difference between Machine Learning and non-Machine Learning algorithms:

Machine learning algorithms improve when you provide them with more data. Non-machine learning algorithms don’t.

Digit recognizer

A famous and somewhat standard ML beginner project is the handwritten digit recognition project.

As it turns out, this simple task is a really good way to understand the difference between ML and non-ML algorithms. There are a few reasons.

Simple Input

The input to a Digit recognition algorithm is very simple. It is a scanned image of a handwritten digit.

You don’t need a lot of background to understand the problem definition. In contrast, suppose you are trying to identify part-of-speech tags for text, then you first need to understand some basic grammar rules to know what the part-of-speech tags are.

Simple Output

This is a simple classification problem. There are only 10 digits  (0 – 9).

Contrast that with ML based language translation – the output is a lot more complex, and constructing grammatically correct sentences is not very easy even for humans.

Ambiguity

Also, digits are usually easy to visually separate.

Sometimes, it is hard to tell with cats vs dogs 🙂

Another nice thing about the digit classification project is that it is more or less universal. There isn’t any subjective opinion about the correct answer. In contrast, when asked to classify sentences from the abstract of a medical research paper, medical experts themselves did not agree with each other always.

The accuracy of the digit recognition algorithms improved significantly [1] once more data was available to train the algorithm.

If you are familiar with Python programming, I suggest going through this article to understand how to solve the digit recognition problem.

More importantly, it will give you a good idea of the difference between an ML based and a non-ML based approach when creating algorithms.

[1] The paper here describes this, if you are into such geeky stuff.