AI Explained Easily

语言: CN / TW / HK

Understand AI in 5 minutes

How would you explain to a medieval time traveler what cars are? Simply, there is too much knowledge missing: he has no idea of what electricity, fuel, and batteries are. What you can say to him is that a car is a tool that can bring you from point A to point B. Seriously, does he need to know anything else?

Like a car, you do not need to be knowledgeable of the mechanical design and physics of how a combustible engine works to understand its use.

A simple scheme of how you can represent the decision process of an AI: features, labels, and rules. Later on, it will be explained in detail.

Problem

The only reason you might want to know what AI is is to know what problems you can solve with it.

What do I mean by the term problem? If you wish you use AI, you need data. What you want to extract from the data is essentially a mathematical problem that you have to solve. Some examples of problems are:

For example, if you have a dataset (a dataset is essentially an Excel table with data) on Covid-19, you might want to forecast how many new cases there will be tomorrow in the World.

If you have a dataset of millions of emails you want to divide the spam from the non-spam , and if you have a dataset of road signs, you want the computer to be able to recognize them by classifying them into different categories .

Here is a list of common problems that AI can solve for you. If you happen to have any dataset with information and know what problem you want to solve, with the proper tools you can create an AI with a personal computer (even on Google Drive) for FREE . This gives you an idea of how available this technology is in 2020.

A list of the most common problems you can solve using AI

These are the most common categories of statistical models in Machine Learning. You can choose which one to use to solve your problem. (the two missing statistical models are called Association and Dimensionality reduction, but they are not so practical, therefore will not be covered in this article).

Why so many examples?

Of course, the problems you can solve with AI are not limited to the ones listed here, but this is a collection of the most practical and common problems.

If I was going to list one example only, it is unlikely you would have understood its practical purposes. With so many examples you can see the practical applications in different fields. If you are managing a company and you wish to make good use of its data, you have many options: forecasting future earnings, understanding your target in more detail…

AI = Machine Learning

To minimize its complexity, you can think of AI as a collection of statistical models that can be applied to data. This vast collection is called Machine Learning.

We can choose from many statistical models

Machine learning is also divided into two subcategories: Supervised and Unsupervised. Each category contains plenty of statistical models. Given our problem, we just have to choose the statistical approach.

Ex.

Given a problem we choose the best statistical model

We stated our problem: we wish to predict how many positive new Covid-19 cases there will in our country tomorrow. Among the plethora of statistical approaches, we choose to use the first graph, which is a regression approach. (It takes a data analyst to select the best model with the best parameters, that is the reason why in this article you won’t find any information on how we choose one model from another).

The Role of data

Because there are two categories of problems that can be solved with AI, we will look into Supervised Learning first. The following example shows a tabular data, a simple Excel file: a dataset.

To use AI, you need data. The more data, the better.

Above we have the data of all 627 Titanic passengers. Problem: Can we build an AI that can predict who can survive the ‘Next Titanic’?

We know the problem. Now we have to choose the statistical model:

Machine Learning > Supervised Learning > Classification

Classification is the statistical model we have been choosing

Features and Labels

The meta-process that is common across all Artificial Intelligences is the following; that is all you need to understand. We isolate the data into two parts:

Features

The columns that allow us to predict the data. In this case: sex, age, siblings, parch, fare, class, deck, embark_town, alone

Labels

The columns that we want to predict. In this case, the column ‘survived’.

Problem: given the features of a new passenger, n. 628 (female, 24 years old, 1 sibling, parch 0, 75.05 USD fare, First class, deck D, embarked from Cherbourg, not alone): will she survive?

Now that we have features and labels, the AI will come up with some rules.

Training the AI: discovering the rules

With the use of the algorithms we have chosen (our statistical model), rules have been discovered. We can use the same rules to solve our problem and find out who will be able to survive the Titanic based on his data (sex, age, siblings…).

Making predictions: A Second Titanic

Now, let us assume there is a second Titanic. We still do not know who has survived yet (labels ‘survived’ are missing); We have features, and because our AI has already been trained, we also have rules.

We do not yet know who will survive from Titanic 2

Given features and rules, we can predict the labels: who will survive the Titanic 2.

Summary

This is an outline of the entire process:

Is that it?

What I have been showing you is one of the three paradigms of AI: Supervised Learning.

When you are lucky enough to have both features and labels you can you a Supervised learning approach to solve it. If you only find yourself with features (a collection of images of cats and dogs, for example), that is called Unsupervised learning: you will use different statistical models.

***The third paradigm is called Reinforcement Learning, but that is not a practical thing to know at the moment.

Why so many algorithms?

When you increase in complexity, you can see that each AI approach has a myriad of different algorithms. To solve your specific problem, you need a specific algorithm.

What is the difference between AI and Statistics?

You may be curious about what is the actual difference between AI and statistics at this point. With statistics, you can make predictions, the same with Artificial Intelligence. Is there any difference?

Simply put: Artificial Intelligence is a collection of statistical tools that are used to make predictions.

What I left out in this Article

The purpose of this article was to give you an idea of how you could use an AI if you have data to analyze. Deep Learning, one very powerful kind of AI, was not mentioned because of simplification purposes. At the same time, I left out other detailed procedures of AI training, including feature selection and testing.

Therefore, consider the content of this article a general map of the principles required to understand an AI.

分享到: