What Exact Machine Learning?
Brach of computer since. different types of machine learning algorithms within machine learning so before we get started let's actually talk about what machine learning is I think we can loosely define machine learning as either an algorithm or program that you built which can solve a very specific problem exactly like how a human would now there's actually a lot of different type of machine learning algorithms but there's four overarching family types as you can say or for overarching categories of these machine learning algorithms.
• Supervised learning
• Unsupervised learning
• Self-supervised learning
• Reinforcement learning
Now obviously as more research is being done into machine learning we're probably going to see a lot more different types of categories within this but for now these are the four main categories that are out there now to be honest most of the machine learning that we know or we have personally used they all fall under.
Supervised learning
supervised learning in fact most of the machine learning solutions which are out there they are all supervised learning and the all all the other three uh types of machine learning algorithms they are actually pretty new in fact uh reinforcement learning is probably the newest type of machine learning algorithm out there uh Google has actually built a program which plays the game go even better than humans and those that program actually makes use of reinforcement learning in the future we're probably going to see a lot more use for reinforcement learning probably in self-driving cars and whatnot but for now most of the machine learning that is being used they use uh supervised learning so let's actually start off and discuss what supervised learning is so supervised learning as the name suggests makes use of a lot of human supervision so for example a supervised learning model would have a test case and a training case provided so the training cases is when you know uh the inputs and outputs are already mapped so the model looks at the training case uh scenario it is trained on the training case scenario and so it's able to learn from the training case which is probably uh you know classified by a human or some sort of code so what happens is after it has been trained by the training cases it then looks at the inputs from the test case and it tries to map it to the outputs in the test cases so in supervised learning what is really important to mention is that the type of outputs it's already provided all the possibilities of outputs that are possible they're already provided it's already known and also there is a training case for this supervised learning model so that is the biggest difference between supervised learning and other types of learning out there now let's let me actually break it down by actually giving you guys an example for example you have a uh you have a data set with images of cats and dogs and you want to train a supervised learning model which is going to be able to look at images of cats and dogs and dogs and correctly identify if the image that is looking at is either a cat or a dog so to do so what you would do is you would actually break your data set into both a training case and a test case a test data set so in the training data set what happens is you've already labeled the images so if it's an image of a cat it's already labeled as you know cat dot png and if it's an image of a dog it's already labeled as dog dog png so that the train so the supervised learning model is able to look at these images and know that okay it's looking at an image of a cat and a dog and what not so it's trained on that and what happens is it starts to learn from this training data set and then it's going to look at the remaining half of the images in the test data set which are unlabeled so you haven't labeled them it's going to look at those images and either predict if the the image that is looking at is that of a cat or a dog so the outputs are known so that this algorithm can only predict either if it's a cat or a dog it's not going to suddenly say that okay it's actually a frog you know so the outputs are known it has a training data set it's being trained and it's looking at the input and you know classifying if it's either a cat or a dog so this is actually a very basic example of supervised learning some of the most uh common uses of supervised learning is actually image recognition object detection uh classification issues a lot of classification algorithms they fall into supervised learning so you can have a multiple of different classes obviously you can have more than just cats and dogs you can have numerous different classes but they all fall under supervised learning.
Unsupervised Learning.
unsupervised learning now as the name suggests it's unsupervised so what exactly does that mean well in supervised learning we knew we had a set of well-known outputs that we had to look out for and also we had a training aspect to our model so when it comes to unsupervised learning we don't have any of the so we don't exactly know the exact type of outputs we're looking for and also there's no training aspect to unsupervised learning so what what exactly does it do so basically it looks at a bunch of data that you have and actually tries to identify patterns so unsupervised learning is able to identify and pull out patterns from your data that you might be looking for so it's really useful especially when it comes to data analytics and interestingly enough unsupervised learning is often like the first step before you actually move on to supervised learning so what exactly does that mean so unsupervised learning is actually used in data pre-processing before you actually use maybe a supervised learning algorithm so unsupervised learning is actually able to process your data better so you can easily feed it into maybe a supervised learning algorithm and it's much better to work with to give you guys an example of unsupervised learning algorithms some of the most famous ones include clustering and reduction of dimensions what exactly is reduction dimensionality reduction well let's say you're looking at a really big data set about houses so it includes the price the square footage maybe even the facing the location and maybe 100 more types of you know types of data about houses you don't exactly need all of them so for example if you're just trying to predict housing prices you probably don't really need a lot of those type of data so dimensionally dimensionality reduction what it does is it actually removes uh these type of like
data points which you might actually not need and that that helps to actually narrow down your data better to feed it probably into a supervised learning algorithm and make it more efficient so unsupervised learning is often usually the first step and it's used in data preprocessing before you actually move on to maybe supervised learning.
Self-supervised learning.
self-supervised learning which is the third form of machine learning algorithms what happens is there is no human involvement involved so you're not exactly labeling the data for the model but the model actually is able to label the data itself and then use those labels as outputs and able to do predictions based off of that it's very similar to supervised learning so it's really easy to confuse but the the aspect where in supervised learning you're actually providing the outputs this in self-supervised learning is what the model is able to do the last type of machine learning algorithms that.
Reinforcement learning
we're going to look at is reinforcement learning now reinforcement learning is probably the most newest type of these machine learning algorithms out there and there's a lot of research that is being done on them there's a lot of future prospects and definitely a lot a lot of opportunities for use use cases for reinforcement learning in the future but as of now it's mainly used in a lot of gaming situations so for example as I was saying you know Google has actually built a program which plays the game goal and that is based off of reinforcement learning now obviously go is not that complicated of a game but let's compare it to a more complicated game like data so open-air has actually built a bot based off of reinforcement learning which plays the game data really well in fact it has actually went against one of the best teams in the world in data 4g and has actually beat them as well so this is really uh definitely a huge win for machine learning and obviously reinforcement learning in the future we can definitely hope to see reinforcement learning not just in the gaming sectors but we can definitely hope to see them in more wide use case scenarios for example self-driving cars so the way that reinforcement learning works is is that there is an agent involved so for example the reinforcement learning model has an agent which is always constantly looking at gaining information from the environment that it's in and then making decisions based off of that to reach the goal that it's supposed to so when it comes to playing a game for a reinforcement learning model what it does is the agent is constantly making decisions in order to reach the highest possible score that it should so that's basically how reinforcement learning works on the topic of using reinforcement learning for self-driving cars amazon has actually released this thing called speed racer I’ve what speed racer is that it's essentially a virtual racing situation that and you can actually program your racing car using a reinforcement learning model and you can actually compete with a bunch of people all over the world so based off of this we know there's a definitely huge potential for a lot of these machine learning models obviously majority of the machine learning solutions that we see these days they are based off of supervised learning but i think there is definitely a huge shift towards reinforcement learning and I think we can definitely expect to see that in the next five years or so well guys I hope this blog was helpful in identifying the different types of machine learning.
No comments: