Deep learning: what is it and how is it related to machine learning?

What is Deep Learning ?

Deep Learning is a subset of techniques that were born around the 2000s as a result of Machine Learning . For this reason, we should classify it as one of its branches, being in turn part of computer science.

These systems are more autonomous than their older siblings, although their structure is also considerably more complex. This gives them a clear advantage when performing different types of tasks where they perform the same or better work than other systems with Machine Learning algorithms .

Also, there are other works where Deep Learning stands out over its predecessor. One of the most notorious cases is AlphaGo- style Artificial Intelligence , the Intelligence of Google capable of defeating the world champion of Go .

Maybe it sounds a little Chinese to you, but Go is a very famous game and, also, very demanding. To put it in context, mathematicians emphatically claim that this hobby is considerably more complex than chess.

On the other hand, Deep Learning is closely related to Big Data, since these great sources of information can be used to learn and consolidate experience. Furthermore, thanks to the situation we are in, the environment for the proliferation and development of this technology is perfect for three key points:

The great accumulation of data, since with the tools we have today, data can be obtained and stored from almost anyone. The degree of technology we are in, since the components are good to collectively offer considerable power. The desire of companies to improve their methodologies, since, taking advantage of the two previous points, more and more companies are betting on Artificial Intelligence . If your company has stored data from thousands of customers and technology gives you the opportunity to learn from them and use it, it is a safe bet.

The structure of Deep Learning

Despite having a development quite similar to Machine Learning , this set of algorithms have some nuclear differences. The most important is probably its internal structure, that is, the code that makes up its algorithm.

General idea about Deep Learning

As you can see in the image, Deep Learning is closely related to neural networks. This concept is not new, but it has not been with us for a long time, so you may not know it.

To simplify it, we could define a neural network as a set of algorithms (each called a layer) that treat and transmit information. Each layer receives input values and returns output ones, and as it passes through the entire network, a final resulting value is returned. All this, happens sequentially, normally, where each layer has a different weight, depending on the desired result.

Here we show you a short video (in English) about Artificial Intelligence learning to play Super Mario World :

And you may be wondering, "Why is all this method so intricate?" . Certainly Deep Learning still belongs to what we call Weak Artificial Intelligence , but it is possibly the first step towards strong.

This methodology is loosely inspired by how a brain works. Similar to what we see in the "physical world" , systems form layers and each layer works in a similar way to a neuron. In this way, the layers relate to each other, share information and the most important thing is that everything is done autonomously.

Very simplified scheme of how Deep Learning works

Following this rule, the most complete Intelligences are, normally, those that have more layers and more sophisticated algorithms.

How does Artificial Intelligence work with this algorithm?

If you have seen our previous articles on the subject, you will have already seen this gif. Here you can see our article on Artificial Intelligence and here you can read a little about Machine Learning .

but we will show you one last time.

This image reflects well and very simply how an Intelligence using neural networks would work. As you can see, his job is simple: classify images and learn to detect dogs in the different photos that are passed to him.

Each image begins by entering the input feed, that is, the Input Layer where the first calculations would already begin. The results obtained would be shared to the second layer or neuron and, evidently, it is informed which neuron has made this calculation. This process is repeated as many times as layers our system has until we reach the last one.

The last neuron is named as the Output Layer and is the one that, in this example, shows the result. In other cases, the Output Layer ends up performing the calculated action. Also, if we put into the formula having to act as fast as possible (like in video games) , the result should be almost instantaneous. However, thanks to the technological point we are at, this is already possible.

One of the clearest examples of this is AlphaStar Artificial Intelligence, another creation of Google itself.

Google Deepmind Artificial Intelligence

We have told you about AlphaGo , an AI capable of fighting against the best Go players in the world. However, this one has younger siblings capable of achieving some pretty impressive milestones.

AlphaZero

This Intelligence learned in just 24 hours a superhuman level of chess, shoji and go with which he won several famous players. Also, in the list of defeated opponents he also pointed to the AlphaGo Zero version of 3 days of experience, something really incredible. Here the speed of learning of this Artificial Intelligence comes out .

Most impressive of all, the team did not have access to learning books or databases, so all of their tactics were learned with practice.

In another of his encounters, he faced Stockfish , a veteran automated open source program that plays chess. However, in just four hours it was dominated by AlphaZero.

It should be noted that while this first calculates about 70 million movements, AlphaZero, in chess, only takes into account 80 thousand different exits. The difference in predictions was offset by much better judgment of what would be promising plays.

With demonstrations of force like this we can see the power of the new Artificial Intelligence .

AlphaStar

On the other hand, AlphaStar is an AI that, today, is capable of playing RTS Starcraft II (Real Time Strategy, in Spanish).

At the time of its demo, AlphaStar fought several professional players in the middle winning ten games in a row and only losing the last one.

Unlike chess or go, Starcraft II is a real-time matchup, so every second you have to be doing things. Due to this, we can glimpse that current technology is capable of maintaining these frenetic rhythms of calculation and decision.

As for the preparation of the Intelligence , for the dates of the live test he had around 200 years of experience training only with protos (one of the available races) . It was also trained so that it could only perform actions if it had the camera physically on the unit, thus assimilating more to how a person would play.

However, despite having these handicaps, AlphaStar managed to beat most of their encounters using an abandoned tactic on the competitive side of the game. One point to note is that AlphaStar usually keeps APMs (Actions Per Minute) low, so its decisions are very efficient.

Average actions per minute performed by the AI and by a professional player

However, when the situation calls for it, he demonstrates superhuman control of units literally by easily breaking the counter.

Here you can see one of his demos in full:

The future of Artificial Intelligence

We have already talked about this topic, so we will not repeat the same talk too much. What should be highlighted are the possible futures that await Deep Learning .

According to Andrew Yan-Tak Ng, a well-known expert in Artificial Intelligence, Deep Learning is a good step towards the Intelligence of the future. Unlike other teaching methods, this one is considerably more efficient as we increase the data sample.

WE RECOMMEND YOU BABAHU X1: The AI toothbrush is now available

The next slide belongs to his presentation "What Information Scientists Should Know About Deep Learning . " If you are interested, you can see it at this link.

Not in vain, the development of technology has not stopped. Every year we will have more powerful components, so we will have more and more patio to test. As happened with old AIs and Machine Learning, new algorithms, methodologies and systems will appear and replace today's innovative Deep Learning .

Also, as you can imagine, the future is tackled by semi-intelligent machines.

As we pointed out in other articles, most electronic devices will have (some already incorporate them) Support intelligence . A very notable case is that of the Intelligences that help to take better quality photos.

However, a point where this technology may flourish for most users is the IoT (Internet of Things, in Spanish).

The Internet of Things

This term has more and more weight in the conferences of technology and computing and seeks to consolidate itself now that we have the means.

The idea is that household appliances, electrical appliances and others are identifiable objects, they can communicate with each other and, in addition, be controlled with a device. In this way we can have a count of what objects exist in a place, where they are, interact with them and all this from the mobile. Likewise, the objects could also interact with each other and if for example a food expires, perhaps the refrigerator would be able to tell you when you open it.

On the other hand, Artificial Intelligence should be able to monitor the status and performance of household appliances. With this, you could establish an electricity plan and optimize the energy used.

However, a relevant point that remains for us to improve would be Internet security. It is something that still does not seem to suffer much harassment, but we all know that it will be essential if we want it to be a safe service.

It is a somewhat abstract idea, but as it invades our lives, you will become familiar.

The importance of new technologies and Deep Learning

It is inescapable to think that computing and Artificial Intelligence are going to shape much of the future that awaits us. Therefore, it is important to always be half aware of what is happening in the world governed by bits.

With that spirit in mind, we can already see how different degrees, courses and degrees appear that teach these topics in depth. For example, some data engineering have appeared, other degrees on Big Data and, clearly, courses in Deep Learning and Artificial Intelligence .

For that same reason, we urge you to investigate the subject. The Internet , with its pluses and minuses, is not yet autonomous, nor perfect, nor really secure, but it is an almost unlimited source of knowledge. With any luck, you will find a place to learn and you can embark on a new language, or rather, a new world.

Since Machine Learning is a slightly lighter discipline , there are programs that allow you to mess around with the data a bit. If you are interested in learning a little more about the subject and checking for yourself / the limits of this technology, you can visit IBM Watson Developer Cloud or Amazon Machine Learning. We warn you: you will have to create an account and it will not be an easy way to learn, but perhaps one day it will help you achieve great goals.

Beyond here is the world of ideas, so everything is in your hands. And to you, what do you think of the new technologies related to Artificial Intelligence? What other Deep Learning applications do you know or would like to see? Share your ideas in the box below.

Source Business Blog Think BigXatakaMachine Learning Mastery