# Deep Learning for Tech Product Managers by fmr PlayStation Sr PM, Jonathan Farrow

## The Perceptron algorithm.

This algorithm is the simplest neural network. Once you understand this algorithm, you can create more complex algorithms and neural networks. It really is a fudge.
To have an understanding of deep learning and the kind of models and data in general, we focus on user experience, business, and technology.
Look at the data and determine if this algorithm can be applied to it. This applies to most algorithms, they are suitable for a certain type of data.

This algorithm is the simplest neural network. Once you understand this algorithm, you can create more complex algorithms and neural networks. It really is a fudge.
To have an understanding of deep learning and the kind of models and data in general, we focus on user experience, business, and technology.
Look at the data and determine if this algorithm can be applied to it. This applies to most algorithms, they are suitable for a certain type of data. The perception algorithm is suitable for data that is a kind of binary output.
If you look at this kind of data here, there is data that we consider a miss, a failure. And it can be anything. It can be sold houses - houses were not sold, a passing score is a failing score. Here is the result of 1 and 0, there are two kinds of classification. Now this is suitable for the perception algorithm.
Let's draw a line, everything above the line is a pass, and everything below the line is a failure. As we enter new data points, the algorithm can predict whether they will be above or below the line.

How to draw this line?

Where Y is the target, predicted variable. X is the predictor, and M and B are the model parameters, M is the slope and the larger it is, the steeper the line and B intercept because this is the value at which the line crosses the y-axis.

Add as many variables to the equation as you have predictions.

When b=0, it's going to cross 0, the rate of change picks up quite dramatically. Everything doubles here.
Where b=500, you can see how these parameters affect the line.
This is a good introduction to show what the parameters actually do.

How does all this fit into deep learning?

Then there will be something called a step function, which will then convert that to 1 or 0, and the result will make it clear whether it passed or not.
What is the perception algorithm? The concept is very simple. First, it randomizes these parameters. First you need to choose some random values ​​​​and say "I'll draw this line." Usually this will have 50/50 accuracy. So when I state any of these points, maybe something will turn out right. It might go wrong. We have a random predicted model that needs to be forced to change somehow. We have to say "no, this line is wrong, please change it".

For every point this is misclassified it will inform the it needs to change.

Perceptron Error correction works by adding or subtracting the error value from the weights and bias.

We multiply by 0.01 with the predictor and subtract it from the weight. We've made the change much smaller, which means the line won't change as drastically. This is good, because we need to make small changes and save the data until we get good results. If there are large fluctuations, we can get stuck on those indicators where the line constantly moves up and down. That's why we introduce a small learning rate. We do this for every point we transmit over such a simple network.

The line that we have now will make a good forecast for us. We can see how good this prediction is by looking at the root mean squared error. And when it gets small enough, we can say that our model is good.
We take all the errors, square them and subtract from the average. You subtract each point from the average and then divide that by the number of points you have.
Now the model is good, but small enough to make a good enough prediction for us. Thus, we can evaluate whether our model is good enough and when to stop.

It can be multidimensional data, multidimensional plan. We will build a model on this. Now we are going to apply some bug fixes to this. We will change the parameters and see if the model is good, if our error function is small enough. This is the beginning of the process. We can start making predictions and say "this is above the line" or "this is below the line". And this is the very beginning of deep learning.
Now we can expand this one more level.

This diagram shows that you can do this with a multiple perception model.

As you start adding more perceptions, your models become more and more complex. And you can start predicting bigger outcomes and more complex patterns.

## Summary

• Perceptron algorithm works on binary data

• We will randomize the parameters

• The algorithm will try use an error function to learn the parameters

• It will output a binary 1 or 0 for its prediction.

`Jonathan Farrow, fmr PlayStation Sr PM`