Introduction to Machine Learning
The world is filled with things that we are able to react to and understand without much thought for example consider a stop sign that’s partially covered by Snow it’s still a stop sign or a chair that’s five times bigger than usual but it’s still a place to sit but for computers who don’t have the benefit of growing up and learning the nuances of these objects the world is often much more messy and complicated. For the first topic we will start with a video called making sense of a messy world where Google engineers and researchers discuss how machine learning is beginning to make computers and many of the things that we use them for such as maps search recommended videos translations and so much better. So you have heard a lot about machine learning or ml lets start with the definition what is ML here is the definition I like to use ML a way to get predictive insights from data to make repeated decisions offer you do this using algorithms that are relatively general and applicable to a wide variety of data sets think of a company in how they user data today perhaps they have a dashboard that business analysts and decision-makers view on a daily basis or report that’s read on a monthly basis this is an example of a backward-looking use of data looking at historical data to create reports and dashboards people tend to mean when they talk of BI or business intelligence a lot of data analytics is backward-looking nothing wrong when istead you use ml or machine learning to generate forward-looking or predictive insights of course the point of looking at historical data might be to make those decisions perhaps business analysts examine the data and they suggest new policies or rules they can suggest for example that’s possible to raise the price of a product in a certain region now that business analyst is making a predictive insight but is that scalable can the business analyst make such a decision for every single product in every single region and can they dynamically adjust the price every second now here is where the computers get involved in order to make decisions around predictive insights repeatable you need a ml you need a computer program to derive those insights us for you. So ml is about making predictive insights from data many the minute I it’s like scaling up BI and decision-making the other part of the machinery definition is around the use of standard algorithms ML uses standard algorithms to solve what looks like seemingly different problems normally when we think of computers if you don’t we think our programs that do different things for example the staff were they used to file your taxes is a very different place of where they used to get directions home when you’re driving. Machine learning is a little different you use the same software under the hood, that’s where we mean when we say ml uses standard algorithms but you can train that software to do very different things you can train the software to estimate amount of taxes that you owe or training that same software to estimate the amount of time it will take to get you home. The ML software yout train on your specific use case is called a model so you now have a model that can estimate your taxes or the time to get you home we use the term model because it’s an approximation it’s a model of reality. For example we have given the computer lots of historical data on drive times and New York city and it will learn the relationships and the data traffic powder and season our time of day impact to predict today’s commute time home whatever the domain ml modelling requires lots of training examples. We will train the model to estimate tax by showing it many many many examples of prior year tax returns, we will train the model to estimate trip duration by showing it many many many different journeys so the first stage of ml is to train ml model with lots of good examples. An example consists of an input in the correct answer for that input that’s called the label in the case of structured data that is rows and columns of data an input can simply be a single row of data in unstructured data like images an input can be a single image say like a cloud that you want to classify is this as a rain cloud or is this not. Now imagine you work for a manufacturing company you want to train the machinery model to detect defects in these parts before they get assembled into the final product for users. Well you could start by collecting a data set of the images for these parts. Some of these parts of the good some of these parts couls be fractured or broken up and for each image you will asign the corresponding label thats the right answer broken or not broken apart and then use this certain example as training data for your model after you train the model you can then used to predict the label of images that it has never seen before. Learn from the past predict for the future here your input for the training model is an image of the part because the model has already been trained its correctly able to predict at this party is in good condition note that the image here is different from the ones used in our training examples but it still works because the ml model has a generalized and hasn’t memorize the training did of the specific examples that you shall know it is learned a more general idea of what a good looking part with a good condition for that part looks like so why do we say these algorithms are standard well the algorithms exist independently of your of your use case even though detecting manufacturing defects and parts and in those images and detecting something like diseased leaves and tree images are two very different use cases the same algorithm in image classification network works for both. Similarly the worst rated albums for predicting the future value of a Time series data set or transcribe human speech to text rezna is a standard algorithm for image classification know it’s not crucial understand how an image classification algorithm works only that it’s the algorithm that you should use if you need to classify images of Arts when you use the same algorithm and different data sets there a different features or inputs relative to the different use cases and you can see them represented visually here. You might be asking yourself isn’t the logic different you can possibly use the same rules for identifying defects and manufacturing that you do when identifying different types of leaves you right the logic is different but ml doesn’t use logical if then rules. The image classification network isn’t like that out of rules if this then that but a function that learns how to distinguish between categories of images so even though we start with the same standard algorithm after training the trained model that classifies leaves is different from the trained model that classifies manufacturing parts and guess what you can actually re use the same code for the other use cases focused on the same kind of task so when our example we’re identifying manufacturing defects but the high-level tests with classifying images you can reuse the same code for another image classification problem like finding examples of your products in photos posted on social media however you still have to train it separately for each use case. The main thing to know is that for machine learning your model will only be as good as your data and more often than not you you are a lot of data from machinery for our example that we talked about you need a large dataset of historical examples of both rejected parts and parts in good condition in order to train a model to categorise the part that is defective or not the basic reason why a ml models need high-quality training data is because they don’t have human general knowledge like we do data is the only thing that they have access to to learn from