What is the difference between artificial intelligence, machine learning and deep learning?
"Artificial intelligence is that the technology of the longer term ," "artificial intelligence is scieence fiction," "artificial intelligence is already a part of our daily lives." It depends on what you're saying by pointing to the face.
For example, I will explain how DeepMind won when the program "AlphaGo" developed by Google DeepMind defeated Korean professional Go player Lee Se-dol in a game of Go this year. Therefore, the terms "AI", "machine learning", and "deep learning" have been widely covered in the media. All three are part of the reason AlphaGo defeated Lee Sedol Go, but they are not the same.
When considering the relationship, it is easiest to express it as concentric circles. First of all, the first idea that came up was "AI". This is the most comprehensive concept. Next, "machine learning" evolved, and finally "deep learning" appeared. Deep learning, which drives the rapid growth of AI today, is included in AI and machine learning.
The history of AI that has repeatedly risen and fallen
AI is the product of human creativity. Since several computer scientists gathered under the name of AI at the Dartmouth Conference held in 1956 and established the field of AI, full-scale research has been carried out at various laboratories. .. Over the next few decades, AI was welcomed as "the key to a very bright future in human civilization," and seemed useless as "a reckless concept of technology-vulnerable scientists." It was sometimes treated as. In short, until 2012, the rise and fall of the rise and fall was repeated.
But in the last few years (especially since 2015), AI has seen rapid growth. Most of the reason is due to the widespread use of GPUs for faster, cheaper and more powerful parallel processing than ever before. In addition, two important elements are simultaneously appearing: "virtually infinite storage" and "a huge amount of data of all kinds (or the movement of such big data as a whole) such as images, texts, transactions, mapping data, etc." It's also related to what you did.
Let's take a look at the history of computer scientists in the transition from the slump until 2012 to the boom that created the various fields of application that hundreds of millions of people use every day.
Artificial intelligence – Human intelligence reproduced by machines
 |
| "King Me": Computer programs that played checkered games were one of the earliest examples of artificial intelligence. |
It caused a wave of excitement in the early days of AI in the 1950s.
Back in the summer of 1956, when the Dartmouth Workshop was held, the dreams of the AI pioneers who proposed the conference were "building a complex machine with the same characteristics as human intelligence that can be realized by emerging computers. To do. " We think of this as a "general AI" -a marvelous machine that has all (or better) human senses and all judgment and thinks like a human being. Thing. Machines like this have been portrayed endlessly in various films (sometimes as "friends" like Star Wars' "C-3PO" and sometimes as "enemy" like "Terminator"). .. There are good reasons why general-purpose AI machines still appear in movies and science fiction novels. Because humanity has not (at least not yet) realized them.
What humanity has achieved so far applies to the concept of "specialized AI" (Narrow AI). It's a technology that can do human-like (or better) processing of a particular task. Examples of specialized AI include image classification on services such as Pinterest and face recognition on Facebook.
These are examples of specialized AI that have been put to practical use, and such technologies reproduce certain aspects of human intelligence. So how is that intelligence realized and where does it come from? So the story moves on to the next circle, "machine learning."
Machine Learning – Approach to Realizing Artificial Intelligence
In essence, "machine learning" is a method that uses algorithms to analyze data about specific events in the world, learn from the results, and make decisions and predictions. That is, machine learning provides the ability to learn large amounts of data and how to perform a task, rather than manually coding software routines along clear steps to complete a particular task. The machine is "trained" based on the algorithm.
Machine learning was born directly from the spirit of early AI researchers, but over the years, its algorithmic approaches include decision tree learning, inductive logic programming, clustering, reinforcement learning, and Bayesian networks. The technique of doing was adopted. As you know, none of them achieved the ultimate goal of general-purpose AI, and even specialized AI was rarely achieved with early machine learning approaches.
After all, one of the best areas of application for machine learning has been "computer vision" for many years, but it still required a lot of manual coding to complete the job. For example, when identifying a stop road sign, an edge detection filter that allows the program to identify the start and end positions of an object, and shape detection to determine whether it is an octagon, called "STOP". I had to manually code various classifiers, such as classifiers for recognizing characters. And from all those classifiers, one was developing an algorithm to understand the meaning of the image, "learn" it, and determine if it was a stop sign.
This technique isn't bad, but it's not surprisingly great. Problems arise, especially if the signs are not clearly visible on a foggy day, or if some of the signs are hidden by trees. Until very recently, computer vision and image detection haven't evolved enough to compete with humans because they are extremely unstable and error-prone.
But time and the right learning algorithms have changed the situation.
Deep Learning – Techniques for implementing machine learning
 |
| "Herding cats": How to extract images of cats from YouTube videos |
It was one of the first groundbreaking demos to demonstrate deep learning.
Another algorithmic approach created by early machine learning researchers is "artificial neural networks," but most of them have disappeared over the decades. Neural networks are inspired by an understanding of the biological workings of the human brain (all interconnects between neurons). However, unlike the biological brain, where all neurons can connect to any other neuron within a particular physical range, artificial neural networks have separate layers, connections, and directions in the transmission of data. It is defined.
For example, suppose you want to extract an image, divide it into many tiles, and input it to the first layer of a neural network. The individual neurons in the first layer pass the data to the second layer. A given task is performed in the second neuron layer, and then the data is passed to the last layer and the same process is repeated until the final output is produced.
Each neuron assigns a weight (probability of correctness for the task being performed) to its input. And the final output is determined by the sum of those weights. Now consider the example of the stop sign above. The attributes of the stop sign image are subdivided and "verified" by neurons (for example, its octagonal shape, bright red color, distinctive text, road sign size, whether it is moving, etc.). The task of neural networks determines if it is a stop sign and presents a "probability vector" (a guess supported by very advanced knowledge) based on the weights. For example, in this example, the system presents an image with a 86% chance of being a stop sign, a 7% chance of being a speed limit sign, a 5% chance of being a kite caught in a tree, and so on. Will do. The network architecture then tells the neural network if it is correct.
Moreover, until very recently, neural networks have been largely shunned by the AI research community, so this example may be too far ahead. A neural network that has existed since the early days of AI, but has hardly produced "intelligence" until now. The problem is that even the most basic neural networks are very computationally demanding and simply not a practical approach. Nonetheless, a small group of heretic studies led by Geoffrey Hinton of the University of Toronto continued their research without giving up, finally succeeding in parallelizing algorithms for supercomputer execution and demonstrating the concept. Did. However, it wasn't until we introduced the GPU into our research that we were able to achieve that goal.
Going back to the stop sign example again, it's very likely that you'll get the wrong answer (in large numbers) while you're optimizing your network, or "training." The only solution is training. Hundreds of thousands or even millions of images are thoroughly optimized for neuron input until they can give the correct answer almost every time, whether foggy or sunny or rainy. Need to load. By doing so, the neuron network will be able to self-teach the appearance of the stop sign. This is true for Facebook's mother's face, or for the cat Andrew Ng succeeded at Google in 2012.
Eun's breakthrough approach is to use such neural networks to essentially significantly scale them, add layers and neurons, and process huge amounts of data in the system. It's about training. In the case of Eun, it was 10 million YouTube video images. He incorporated "deep" into deep learning, which represents all layers of a neural network.
Currently, in some scenarios, image recognition by machines trained using deep learning is beyond human capabilities. The range ranges from cats to identifying tumors and cancer clues in the blood on MRI scans. In addition, Google's AlphaGo learned Go, played against AlphaGo itself over and over again, and trained to optimize its neural network.
Deep learning brings a bright future to AI
Deep learning has enabled many practical applications in machine learning and, by extension, in the AI field as a whole. Deep learning allows you to categorize your tasks in a way that allows (or can be expected) all sorts of assistance from your machine. Unmanned aerial vehicles, more preventive medicine, or more accurate movie proposals have already been put into practical use or are expected to be put into practical use in the future. AI is a technology of the present and future. With the help of deep learning, AI will be able to get closer to the state of science fiction that humanity has long envisioned. Everyone should be able to have C-3PO as a friend and get their own terminator.
0 Comments