 ## Intuitively Understanding Machine Learning

Artificial Intelligence. The buzzword if the decade. Some people are scared of it, others are scornful. Populist voices angrily express their fear of being automated or worse deny the possibility. Is it denial or misunderstanding?

One thing is for sure: artificial intelligence is misunderstood for good reason. The same mentality from Einstein’s famous quote about quantum mechanics applies: “if you think you understand [artificial intelligence], you don’t understand [artificial intelligence]”.

This is especially true for artificial intelligence that is a “black box”: we actually can’t explain the predictions, but we can measure the accuracy of them.

We can measure the accuracy of a prediction by pretending we don’t know the answer and then comparing the predicted value to the actual value. In order to increase the accuracy of our predicted accuracy, we can split the data up as many different ways we want, training on the one section and testing on the other, ten, twenty, or even a hundred times. This is called cross-validation.

Neural networks are famously hard to model. When I say model them, I mean explain why they come up with the prediction that they come up with. This is because they use something called  automated feature selection. This means that the variables that make the prediction are automatically chosen by the neural network from the data.

Each one of these ‘features’ is called a perceptron, and they can be layered, each layer and link between representing the relationship between the features/perceptrons.

These are exactly like the associations that we have baked into our brains, and that’s the analogy to our brains that the “neural network” represents.

## Slope Intercept Form (Yawn...) for Neural Networks (Yay!)

Remember slope-intercept form from basic algebra?  y=mx+b shows the relationship between y and x, which are both variables. In this equation, b is the constant

*eyes glaze over* – “WAKE UP!”

Okay, so now think really hard. We have two variables that we are trying to show a relationship between. Sound important? If we can model the relationship between two variables, we can make predictions and provide insight into why there is a relationship!

So the m in this equation is called the coefficient and it is important because it quantifies the relationship. It allows us to model the relationship! Things get more complicated though, and then much more complicated (we’ll talk in a bit how to use generalize this idea of slop-intercept form to think about neural networks).

Think about adding more variables. Instead of a relationship between two variables, such as height and weight, now we add other features of a person, such as hair color. Some of them won’t have an impact on what we are trying to predict (height), in which case the coefficient will be zero. So then we have multiple m’s/coefficients.

Now’s the kicker. Take each one of those m’s/coefficients and put another equation inside of it, with it’s own m’s/coefficients. Now consider that maybe there’s a relationship between the m’s/coefficients in that m/coefficient and the m’s/coefficients of the other m’s/coefficients. Mmmm.. M&M’s.

## Neural Networks Simplified for Intuition

There are many different types of neural networks, and one thing that differentiates them from each other is the relationship between the perceptrons, or the m’s. This is the more simple way to understand them that provides the intuition. When one perceptron’s impact on the prediction is visualized, it seems disjointed and random, but together they make the prediction.

A great example is a number recognition system. AI’s are famously good at image recognition. For example, one of the features might be the way that a four has a horizontal line in the middle, and that’s all. The prediction power comes from the way that this relates to other lines, parts of numbers, or other types of features. 