Artificial Intelligence (eBook)

How Advance Machine Learning Will Shape The Future Of Our World

Christina Ahmet (Autor)

eBook Download: EPUB

2018
143 Seiten
Shockwave Publishing (Verlag)
661000013045-0 (EAN)

Infos

Currently, Artificial Intelligence (AI) lives amongst the human population. They reside in smartphones. They help people find content on the internet. They learn the behavior of their owners and put out relevant, interesting content to enhance their owner's experience while they are browsing on the internet.

In this book you will learn all about Artificial Intelligence and how it will affect your life in the near future.

Learn exactly what Artificial Intelligence is

Machine Learning

AI and The Internet Of Things

Opportunities for Artificial Intelligence

Intelligent IoT

and much much more!

Textauszug

Chapter 4: How Machines Learn

Previously, the theories behind machine learning were explained thoroughly. However, to better understand the entire process, it is necessary to illustrate it in real-life events.

On the internet, the algorithms are everywhere. The videos people watch are brought to them by the algorithms. When they click on the video, the algorithms take note. When people go on Facebook, the algorithms decide what people see. When people buy something, they set the price, communicate with their fellow algorithms at the bank to monitor the transaction. The stock market is full of bots trading with each other instantaneously. In fact, this book could be brought to the users by algorithms as well.

Before, humans built algorithmic bots by providing them with instructions that humans can easily explain. If A, then B. However, one could easily see the limitations in manual programming. Certain problems are so vast that it is virtually impossible for humans to program an effective bot. Looking back at the stock market, there are millions of transactions going on every second. How does a person program a bot to know which one is bad? There are countless videos on YouTube. How is it possible to program a bot that can recommend the right video for the viewer fairly consistently? How does one program a bot to determine how much a customer is willing to pay for that airline seat? It is true that bots may not give perfect answers, but at least they do a better job than humans.

Funnily enough, while they can work wonders, no one really knows how they work. Not even their creators know how they function. Even so, many companies are very protective when it comes to how their bots work because they are incredibly valuable employees. How the bots are built is a secret. Whatever their models may be, there are three main categories for machine-learning, which are supervised, unsupervised, and reinforcement learning.

Supervised Learning

A common method for teaching an AI system is by using a large chunk of data and labeled examples. Here, the bots are fed a huge amount of data and then they need to study the data and highlight the features of interest. If the data fed are photos and the bots need to say whether they have a dog or written words, then they will look through the photos and try their very best to understand. Again, neither the human programmer nor the bots themselves know how they think, but at least they work relatively well for a robot. When these bots are trained, which means that they can accurately perform their tasks using the labeled examples, then they should be capable enough of doing the same thing when they are fed new data.

This process of teaching a bot using examples is known as supervised learning. Normally, the task of creating and labeling examples fall into the area of responsibilities by online workers who are generally employed through platforms such as Amazon Mechanical Turk.

Of course, to create a bot that can reliably tell the difference between a cat and a catfish would require a lot of labeled datasets, which can mean millions of images and examples just to teach the bot to do one thing right. Still, in an age of big data and widespread data mining, data is relatively easier to access than it once was a decade ago. There are also training datasets available in huge quantities, and they are still growing, such as Google’s Open Image Dataset, which contains about nine million images. At the same time, YouTube has its own labeled video repository known as YouTube-8M which has the link to over seven million labeled videos. ImageNet, which is one of the earliest databases invaluable to machine-learning, now has more than 14 million categorized images. As such, a company or those who want to design bots might give value to computing power more than having access to the labeled database. In recent years, Generative Adversarial Networks have demonstrated how machine-learning systems, when fed only a small amount of labeled data, can create its own fresh data with which to teach themselves. This method could be tremendously useful and can potentially lead to the rise of semi-supervised learning where bots can learn how to perform a certain task using a significantly smaller amount of labeled data than what is originally needed to train a bot using a supervised learning model.

Unsupervised Learning

On the other hand, unsupervised learning does not need as much human involvement. Here, the algorithms try to find patterns in data and look for similarities that can be used to categorize that data. A perfect example here is by clustering fruits that weigh or look similar or cars with a similar weight or size.

Here, the algorithms are not created in advance to look out for a particular set of data. It just takes a look at the data it is given and tries to group everything in there by its similarities. A perfect example of this is Google News, which groups stories on similar topics on a daily basis.

Reinforcement Learning

The best way to describe reinforcement learning is rewarding a pet with a treat when it successfully performs a trick. Except, the pet is a bot, and the reward is the desired outcome which so happens to also be the trick.

Here, the system tries to maximize a reward based on its input data and continues to learn until the best possible outcome is achieved.

A good example here is by looking at Google DeepMind’s Deep Q-network. It has been used to defeat humans in many classic video games. Here, the bot is given a massive amount of data, from the pixels for each and every game and it attempts to figure out different information such as how to jump so that the character does not fall into the pit. The bot also looks at the score in each game and figures out a way to maximize the score in different circumstances. In the case of Mario Bros., the bot will chart a route to guide Mario through each level to collect all the coins, defeat all the monsters, while doing so in the least amount of time possible.

What these bots can achieve is miraculous, and it seems that the sky is the limit. As long as there is data, enough computing power, and some time, a bot can be taught just about anything. But how do they learn exactly? By studying how such a bot can be built, it is possible to understand how they learn without getting into excruciating details of how each and every neuron works.

Suppose that a bot is needed to sort through a collection of photos and divide them into two categories: three or trees. It is easy for humans, even children, but it is impossible for the human to program a bot to recognize the difference. Humans can describe a tree by saying colors, shape, or material, and a number by its shape. But those are words, and bots cannot understand words. They read and understand algorithms, which is impossible to program. So, how does one build a bot to do this job? Instead of building a bot that can sort through the photos, a bot that builds other bots and a bot that teaches those bots are created. The brain of these bots is vastly simpler, which is something a talented programmer can achieve. Here is where the path of machine-learning really branches out. There are two main models of how to make machines learn: genetic breeding models, and deep learning and recursive neural networks.

Genetic Breeding Model (Or Evolutionary Model)

The builder bot builds bots almost by random, except for some instructions given by the programmer. It just connects the wires and modules in the student bots’ brains, resulting in wildly random bots being sent to the teacher bots.

Of course, the teacher bot cannot tell the difference between a three or a tree either (otherwise, there is no need for a builder bot and a teacher bot). The teacher bot does not teach the student bots. The programmer gives the teacher bots some photos of trees and threes as well as an answer key telling which photo is which. Here, the teacher bot tests the student bots. Because the student bots are initially built by random, they will perform poorly at the test. Then, the results are sent to the builder bot. The best student bots are kept, and the rest are discarded. Then, the builder bot gets back to work by creating copies of the student bots and adding additional changes. Then, the second generation student bots are sent to the teacher bot. The teacher bot hands out the test, corrects them, send the result back, and the builder bot gets back to work again. This cycle continues indefinitely.

Basically, this is the “spray and pray” method of programming, which should not work, yet does. This is partly because, in every generation, the builder bot only keeps the best student and then discards the rest. Moreover, there is not just one student bot. There are countless, and the teacher bot can also have millions of questions as well. This building, testing, discarding cycle continues until the desired result is achieved.

At first, the best bots are just lucky because they have the right algorithms. However, when one combines enough lucky bots and keeps only what works, and then randomly makes a few tweaks, there will be a time when a student bot can do its job reasonably well. While the student bots are copied and changed, the average test score is also raised to improve the bots even further. Keep this up, and a bot will eventually emerge that can tell the difference with great accuracy. But how the student bot does this, nobody knows for sure. Even the student bot itself cannot understand it either. Still, it just works.

However, the bot still has a narrow AI. In this example, the student bot is good at exactly...

Produkt-Details

• Erscheint lt. Verlag: 20.11.2018
• Sprache: englisch
• Themenwelt: Informatik ► Theorie / Studium ► Künstliche Intelligenz / Robotik
• Schlagworte: AI • Artificial Intelligence • Computers • Intelligent IoT • internet of things • machine learning • programming