Machine Learning / Data Analysis Basics Video Lecture - Machine Learning

Hello Explorers and welcome to another video. Today I'm going to talk a little bit about machine learning
classification basics so I'm going to talk a little bit about some methods that are really basic to machine learning
so these are some of the
Different categories that you can use for classification there are more of them
But I chose some of them that I know well and some of them that are really good to know about
when you're talking about machine learning
so first off, I'm going to talk about a few baseline
Machine Learning methods or
Algorithms they are very simple, but good to know about and then I'm going to talk about naive Bayes. Which is a very
Known method and also very naive as the name Implies then we talking about
nearest neighbor Decision tree support vector machines
multi-layer perceptrons and lastly
Conditional Neural networks and these are the kind of networks that you build inside of tensorflow and other machine learning
Frameworks, so let's start off
the Baselines here is the zero R and one R and
Let's take the very easiest thing that you can think of when it comes to Machine learning the zero R
the Zero R
Is actually a method where you say that ok I have 10 different choices for instance?
We are going to try to
Figure out which number is the correct number and then you have this really confident guy that tells you
well
Number six is the correct number and there's nothing to say about it
It's always right
So you run your tests and realize that the e is probably right?
let's say 12 percent of the time, so
That's a baseline. It's not a good one, but it's one of them
and let's say that we have some knowledge about the picture with the
different
Numbers in them and we can say that
This confident guy again says that if there is a line through it
then it will align in the picture then we say that this is a 1
I know, it's the 1 it's nothing but 1 and
He probably be right about
25% of the time
That's another metric. We had take just one single thing into account and that
Is what we base our?
Machine learning on and these are good baselines to see that if you find any
Method that we run that are worse than these two then you have
Probably not a good machine learning technique
Next up we have naive bayes
Nave base is something that takes
Every different aspect of your pictures all your inputs and just put them into a probabilistic
Computation so you can say that if this is correct then give it
a number between 1 and 0 and
Then you take the next one give that number between 1 and 0 and you add those up and divide
Them and then you will get a probabilistic
result that will point to one of your numbers, so it's very naive, but it will take all your
different input data and give you some kind of output which could be good and
It seems like even though this is naive. It was a one of the first that was invented
it's actually correct and at a
alarming rate very good because let's say that you have a really really good
Convolutional Neural Network that can find a solution for something by
99% time the naive bayes could actually be as good as
86% of the time and that could be useful useful in some situations if you want a really fast and easy
solution to your problem
Then we have the k nearest neighbor algorithm, and this is a little bit more complicated, but still
pretty simple you take
your
Specific point that you want to classify, and you look at how much it
Goes into different part of your matrix so you see that
We have a some solutions that are in the different directions as you see in my map here
And then you see how much of the overlap is into the different
Choices, and how close it is to the other answers depending on?
its different
input data
So this could also be a very good
Algorithm and it's usually a little bit better than naive Bayes and
It's a little bit tricky to go into the math around this but this is
The simplest way that I found to describe how it works
So instead of naive bayes that takes everything into account and just gives you a probabilistic
Answer this will weight your answer
Depending on its closest neighbors and see what's the most probable answer
depending on its position in the
computed answered space
And then we have the decision trees these are quite simple one of the most known is C4.5
and it's also implemented in Java called
J48 and it's pretty much that you take your different inputs and
You build a tree around them so you take the most significant of your inputs
That will make most differents to your grass and you take that first and put it to the top in this case
We look at is this a triangle or a square?
How many sides do it has or how many?
Square, how many corners do it have for instance so that could be it one metric?
And then the next stop do you take take if it's rotated or straight?
So this is a way to build a tree of different
answers, and
in that case
easily find Your route down to the correct solution
and then you have a little bit more sophisticated method when you come to a
large tree and it's very hard to build a correct tree
then you can actually use a method called random forests this will create as it the name Implies A
huge amount a forest of trees let's say 500 or 100 trees or a 1000 trees
And in this large forest you have all these trees with different branching and they are
Created a little bit differently weighted so you have different
Paths in all of them and then y ou just ask the forest. What do each tree
Think is the correct solution and then the trees will vote
so each tree get one vote and if you let say have
300 trees to think one thing in 200 trees that thinks another thing then
Mayority Will win
So that's a very interesting method of actually finding out
What you're the correct answer is?
support vector machines
They are a lot more complicated and a really interesting way to actually
cluster and classify your data
because you in this case you will take all your data inputs and you will create a
A
Plane or a Multi Plane or Multi dimensional plane
so for instance if you have just true or false then you could just say it's a
simple
Diagram where you divide your choices with a straight line?
But if you're let's say have three inputs
then you will have to use a
three-dimensional
Space and this will be very much more so you will have a very large
dimensional space if you have very a large amount of inputs so this
Data graph will be a
Very large one if you have a multi set of inputs
So this is another way to cluster things and the method really search for things
to have a wide as a gap as possible when it comes to
Dividing your data now
We are coming into the world of Neural networks
And I have a series on this, so
When I go into and actually build a multi-layer perceptron
so you can look to that series and get a deeper knowledge of how
the Multi-layer perception work and how you can build your own?
but the big part to say about the multi layer perceptron is that you have some inputs and
Then you have a set of hidden layers
That are a set amount of nodes in each layer and each layer has the same
amount of nodes in them and
every input
Has a connection to each node with a weight
so you have a full matrix with all the
Answer all the different weights in it and you will train this
Network to create
Weights that are the correct weights for each input that you have in your training set to get the correct outputs
so for instance if you have an input that has
Some value in in the first
box, and then it goes to the next layer it will actually trigger some of the
nodes in the
first layer of your
Neural Network and the Triggered or
activated nodes
will in turn
activate some nodes in the second hidden layer
with their weights and
those activations will activate some nodes in your third, layer and
so on until you activate
one of the outputs later answers and each output
Will actually get an activation, but you are looking for the strongest signal
Through your network, and that's your answer for this question
So that's a multi-layer perceptron
Next up we have the convolutional neural networks
This is something that you usually build in tensorflow or another network
There is a great Javascript
example online, and I have also
Rewritten it in Java. So I have a series on that how to write your own
layers and Your own
Trainers in Java for convolutional Neural Networks
but the Takeaway about
Convolutional Neural networks is that they will decrease the size of your inputs as I talked about
Before what with the multi layer perceptron
There you had each layer being the same size, but in the convolution on your network you do a
convolution which means that you actually
Find some
nodes that will Trigger and then you do a
subsampling so you will decrease the
space of your
problem and then you do a convolution again decrease your space and then
maybe you do a subsampling again and
the last step of a convolution neural network is
usually a fully connected one with some hidden layers as I described with the perceptron, but
the benefit of using a convolutional neural network is that you can decrease the size and get the same
amount of the result on the other side and sometimes even get better results because you
remove biases and
can have
dropouts in the
Network to remove some of the overfitting that you can get into your network and these
structures can be very complex and very deep so if you're looking at the graph on the
Left hand side here above my head
You have some different methods, and this is a little bit of an old picture
But you see here that it actually gets better for each year that I trained it. So you have the first
competition they had in 2010 they had a
error rate of
28% and then the next year
25% the next Year 16% the next year 11% the
Next year 7 percent and 6 percent and the last year that this graph shows there you have the ResNet
with 3.57% errors in a large amount of the image Data and
the layers has actually
Exploded over the years though the first networks were very shallow
Then you have 8 layers 8 layers 19 layers 22 layers and the last year you have
152 layers deep Neural Network, and I believe that the networks of today is even deeper
because google actually figured out a way to cluster some of their
parts in the Network into units, and then they just cluster those a
bunch after each other so that they have a very deep neural network for some of their
different
Applications, so I hope that this video gave you a small
introduction into the subject of
machine learning and
Gave you some things that you could look up and learn more about when it comes to
specific Data mining techniques, or
Machine learning techniques. I hope you found this interesting. I hope you learned something today and
I hope you subscribe if you haven't done that yet, and
I really hope to see you in the next one