If I see a new out a new observation I can say if this observation relates to this particular logic statements then it belongs to discuss if it doesn’t relate to this particular logic statement then it belongs to another class now you can see already that this can become a very hard problem because yes in the slide I mentioned just about four or five properties four or five properties in the image that could help determine if it’s a cat or it’s a dog now if for any reason this dog or the cat is backing the person that took the picture.
I have access to that picture you can see very rapidly that I can see an eye I can see a nose I can see a tongue I could possibly see an ear now my traditional rules would very quickly fail right so this is what helped a lot of people or helped us move from a more systemic approach of hard-coding everything to a most it to a system where we allow the computer understand what features need to predict in a particular class so for the traditional programming we have a computer we’ll give the computer the data so we’ll give the computer the data the cats and the dogs and then we will give it some rules and tell it if this rule is met send it this image belongs to this class if that rule is not met then send it to another class and then from that our computer will taking new examples and try to give us an output.
This is the problem now the problem is the rules we’re giving it rules on how it’s supposed to understand data and we can see that as the data increases as the data becomes a lot different varieties of images of cats and dogs that are not looking at us that are upside down sleeping we may start to have issues of our rules are not robust enough but can we really really have robust rules how robust is robust enough so that’s why we have machine learning now with machine learning all I can tell the computer is I’m not gonna give you rules I’m just going to give you date and I’ll tell you what each data point is so I give you a picture of a dog.
I tell you that that is a dog I give a picture of a cat and I tell you that’s a cut and then from there I want you to learn the representations that make an image a dog and the representations that make an image a cut and with that develop your own rules right so we teach a computer in using in machine learning we teach a computer how to learn we don’t teach it what to do which is it how to learn so that it then knows what to do right so you can see now it becomes very interesting I don’t have to start I don’t have to spend my time coding if an ear if there’s.
If there is an eye if there is a nose if there is a mouth if their legs if there’s hair if the hair is scanty if the hair is brushed if yeah I don’t have to spend my time coding all of that because when you when you start coding some of those things you’ll see very largely how how large would carry some of those scripts can get and yeah however not to be not to throw away traditional programming traditional programming actually in a lot of cases can help to build the baselines for which machine learning would come on on board as a matter of fact I was speaking yesterday.
I said if your problem cannot be thought of in a traditional programming setting then most likely you cannot do that in a machine learning setting a lot of people just think oh okay every problem since we’re talking about machine learning now we’re just gonna jump on machine learning and just start doing machine learning no if your problem cannot be total as a traditional programming problem it cannot be thought of as a machine learning problem so I just explained now the problem of trying to classify an image as either a card or table I’ve been able to show you how I would write the traditional programming logic for identifying identifying the two classes.
I am now shifting and telling you okay but well since I’ve seen that I can do that with traditional programming I can now do it with machine learning right so what is machine learning classification then first of all we need to understand what learning is before we can understand what the she learning classification is learning is any process like as explained by which system improves performance from experience so I take a lot of I give a computer a lot of data and I tell it learn the rules learn the rules for which I have this data and I have this output the output or the names of the data and the computer learns that and it now has to find a way of improving its learning experience from new data when I after it has learnt the rules.
I then give it some new data that it probably didn’t learn from and tell it oh yeah now you have gone to class this is an exam how do you what image is this and it’s able to tell me based on what I have learned based on how I learned this is a cut or this is a dog so machine learning is the study of algorithms that improve their performance at some task T which experience e so well defined learning task is given by the by de by this three statements improving their performance at some task T which experience e so when do we use machine learning we use machine learning when human expertise does not exist right it’s an experimental system it’s an experimental system.
That has the probability or possibility of of performing really really good or performing really really bad but we know that some of sometimes there might be no expertise or that the the amount of information we have to process is way more way more than a human being can actually sit down or will take too much time to process we then give that kind of kind of in data this kind of data or problems to a machine learning algorithm or machine learning system a combination of systems to try to learn representations within data finding patterns and help achieve what we’re trying to achieve another reason or time when we get to use machine learning is when humans can’t explain the their expertise so I was born and I learned how to speak my language from an early age but I never really went to school to understand how to speak it I never really was sat down to be taught the syntaxes of how my language evolved and grew.
I just knew I knew how to speak my language after some time I could respond and reply to my mom when she’s talking how was I able to do that my brain was able to find patterns in how in the words that were spoken how they were spoken and the syntax is there was spoken so with that it then was able to find it was able to connect and tell me a word to say and how to say it whenever.
I needed to say it in my language I cannot explain that expertise to you I can explain why and how I’m able to understand the language but then yeah if I want to train a system to understand how then do I teach it that or if you see this word convert it to this word if you see this word convert it to this word you can rapidly see that for evil example I will have to learn all the words in evil to be able to translate them to English words so that’s a very very huge problem I have to find a way to give a machine learning algorithm that problem to solve.
So models are based on huge amounts of data when you are in the industry you would find that there is tons turns and turns of data running in terabytes running exabytes as a matter of fact and many of the data is useless to a lot of people because they do not understand the techniques the methods that could draw insights from all of that data so because of how much data we have available as images as texts as emails as audio as tabular data we have to find some kind of system to put all of that data into and find a way of drawing insights or identifying patterns within within them.
So when do we use machine learning classification right machine learning classification is one of the most used methods in industry and I’m sure you you might have or even if you haven’t you may get to see regression and regression is basically trying to use past data to develop a system that once I have new data points I can always know I can always fit so in regression we are not trying to find out what class something belongs to we’re trying to just know okay for any impute impute what could be a continuous valued output so for instance I want to predict the price of a car.
I would give impute features impute features for what they kappa bleep we’ll what it cap ibly is for instance the wheel will base how old the car is the size of the car length and breadth and height possibly the amount of the size of the fuel fuel tank and with all of those I can develop some kind of to predict a real-valued output that’s a continuous output of how much exactly the price would be but in classification we’re not trying to predict how we cannot we cannot try to predict how much the price of a car would be we would want to predict maybe the range of prices that a car could fit into a bucket of prices and say okay for instance.
I would say my classes at the end described this car belongs to the two classes lower than 500000 and above five hundred thousand I would then get features those same features and try to predict what bucket the price of the car should fall into or in regression I will be trying to predict the exact amount that car should should be sold for right so you can see the clear difference between both of them in one I’m trying to classification on the Left I am trying to predict what class something belongs to and in regression I’m trying to predict the continuous valued output for a particular impute right so when do we use machine learning classification machine learning.
Classification being one of the most used we use in recognizing patterns and some of the applications involve identifying object instances of a specific class in a digital image so we can check for facial identities or facial expressions so you can use your device today your mobile device and just look at the screen and it can unlock itself how is that how does that happen it’s basically your system understanding that that’s your face and we have facial expression system systems that now can tell what expression you have on your face if you are sad if you are happy if you are just indifferent and with all of that you can link all of that you can begin linking these systems to recommendation engines or systems where someone is looking at his phone.
It’s right side and the recommendation pops up for a nice musical piece for the person and he’s able to listen to it and gets maybe a little bit light-hearted so yeah facial expressions without hand reach handwriting or spoken words where you try to a machine learning algorithm tries to understand what the letters you are writing using your stylus on a screen you you just write and the machine learning algorithm is able to understand this is what you wrote right so it doesn’t matter if you’re a doctor with your handwriting or whatever you are as long as you can write the machine learning algorithm is able to pick out what you wrote and translate it to maybe.
Whatever you it is that you really need to be to say so yeah in keyboard typing for people who don’t like to use keyboards they can just write and have their handwriting converted for them then we have medical images where we try to identify various various patterns within x-rays brain scans blood samples and try to classify those into different areas of different problems different problems that may exist so for cancer and this is an area medical images is an area where a lot of advanced research is ongoing down into understanding how machine learning algorithms are able to see cancer in patients see other underlying conditions by just looking at scans of those of those images of those people rather then we have generating patterns machine learning algorithms.
Now can generate stuff generate text they can generate text they can generate images they can generate motion right so if you follow the trend of applications so currently you’d see that there’s the new trend of difficult defects where someone say something or record the video doing or saying something and they can rapidly put another person’s face on that person’s image and it looks like it is the path it’s another person that’s talking or doing performing the action this is a very trouble trouble trouble some field as we have seen it boy yeah there’s a lot.