Machine learning is a methodology that enables computers to learn from data and interpret new data, and it is a technology that solves problems through algorithms. Artificial neural networks are models that mimic how the human brain works, and they can solve various problems with high accuracy through perceptrons and deep learning.
Recently, the 4th Industrial Revolution has been a hot topic in Korean society. Some people criticize the term as an abstract slogan without substance, symbolizing the political appropriation of science. In their view, data science, the foundation of the Fourth Industrial Revolution, and artificial intelligence, the cutting edge of it, are mere fantasies. But what is machine learning, the technical basis of artificial intelligence, and what is artificial neural networks, the best-known model of machine learning? What is the relationship between the two, and is the artificial intelligence Big Brother that sci-fi visionaries often talk about a reality?
Machine learning, or machine learning, literally means teaching a machine to learn. The machine is a computer, and the data is real-world data. In short, machine learning is a methodology for making real-world data understandable to a computer using various methods. Once a computer understands the data it’s given, an interesting phenomenon is observed. It can interpret new data when it is fed new data that it hasn’t been given before. It’s like a math student who solves a lot of similar types of problems, and is able to pick the right answer with a high probability, even if it’s the first time they’ve seen it.
The “fancy way” of teaching a machine to learn from data is an algorithm. An algorithm is a collection of steps to solve a given problem using the data provided. For example, consider the problem of identifying the flag of Japan from a set of pictures. Typically, you would follow these steps to find the flag of Japan
1. find a picture that is shaped like a square.
2. Find a picture with a red circle in the center of the rectangle.
3. find a picture with a white background except for the red circle.
These three steps will solve the problem of finding the Japanese flag, even if it’s not always correct, which is called an algorithm. If you can train a computer to do these three steps in a way that it can understand, it will be able to categorize Japanese flags as well. This is called machine learning.
Machine learning is divided into unsupervised learning and supervised learning depending on the type of data to be learned. Supervised learning uses training data for which the correct answer exists, while unsupervised learning uses training data for which there is no correct answer. The Japanese flag classification problem above is a classic example of supervised learning. If the image data provided is the Japanese flag, the answer is either correct or incorrect. On the other hand, if we consider the problem of dividing the given people into groups of people who live close to each other, there is no right answer. Therefore, the results will vary greatly depending on the criteria. This is the problem of unsupervised learning. In technical terms, the “right answer” is called a “label” for the data.
Once the data is labeled and the algorithm to use is chosen, the computer can start learning. Once the training is complete, the computer builds its own framework of how to solve a given problem, which is called a model. An important difference between a model and an algorithm is that a model is a tool created by combining data and algorithms. Consider a picture of a rectangular white background with a small red circle in the center. The simple three-step algorithm above says that this picture should be classified as a Japanese flag. However, if this picture were included in the training data and labeled as incorrect, the resulting model would not classify it as a Japanese flag. In other words, a model is a new set of judgments based on an abstract algorithm with concrete criteria for judgment using data.
There are many different algorithms to solve the same problem, which can result in many different models. In the case of supervised learning, once a model is created, the machine outputs a correct or incorrect answer for new data. The quality of the model is evaluated by how accurate these outputs are compared to the actual labels. For example, if you feed a model 100 new pieces of picture data and 70 outputs match the actual labels, the model has an accuracy of 70%. Typically, the model with the highest accuracy among several models is adopted as the best model to solve the problem. For unsupervised learning, accuracy cannot be determined mathematically, so models are evaluated based on how useful they are when applied in real life. There are criteria for mathematically validating the validity of unsupervised learning, but they are beyond the scope of this article.
If the measure of a model’s validity is accuracy, does it matter what the algorithm is, as long as the model is accurate? Consider the following new algorithm for the Japanese flag identification problem.
1. ask the magician if the picture presented is a Japanese flag.
2. follow the answer.
Usually, the validity of an algorithm is proved mathematically. However, this algorithm is clearly not mathematically provable. If the model is 90% accurate, is this algorithm superior to existing algorithms? Surprisingly, the answer is yes. Of course, this example is very extreme and unrealistic. But in any case, the integrity of the algorithm is of little importance in machine learning. Models that use algorithms that have not been proven mathematically sound are called black-box models. Artificial neural networks, which are currently the most hotly debated topic in machine learning circles, are typical black-box models.
They first appeared in the 1943 paper “A Logical Calculus of Ideas Immanent in Nervous Activity” by Warren McCulloh and Walter Pitts. They were abandoned due to their limited predictive power due to insufficient data, but have made a comeback in the 21st century as storage capacity has increased dramatically. Neural network models were essentially born out of the question of what would happen if computers could think the way the human brain works; in other words, neural networks are an attempt to abstract the way the brain works into an algorithm to solve a given problem.
The human brain is made up of brain cells (neurons), which use neurotransmitters to send and receive information between them. Let’s say you want to send information from neuron A to neuron B. When the neurotransmitter released from A’s end arrives at B, it changes the electrical state of B. If the change is below a certain level, the information transfer is invalidated and nothing happens. However, if it is above a certain level, B will pass the information to other neurons in the same way. This level of change in electrical state that a neuron must cross in order to transmit information is called a “threshold”.
The information transfer process doesn’t necessarily have to be a one-to-one correspondence between two neurons; it can be many-to-many, with one neuron passing information to multiple neurons, or conversely, many-to-one, with each neuron receiving input from at least one other neuron. Neurotransmitters in one neuron may decrease the amount of electrical state change in another neuron, while another neuron may increase it. In any case, the total amount of state change from all of these inputs must exceed the neuron’s threshold for the information to continue to propagate.
Importantly, when a neuron transmits information, it selectively decides what to pass on based on the weights between neurons. For example, when solving a math problem, the cerebrum is responsible for solving the math problem. In this case, information about the math problem is actively transmitted to the neurons in the cerebrum and very little to the neurons in the liver (the liver only processes life-sustaining information). In other words, the neurons in the cerebrum are weighted highly and the neurons in the liver are weighted less. In summary, all human thought and behavior is determined by the total amount of information passed through neurons that have exceeded their thresholds, taking into account their weights.
To understand an artificial neural network, consider a neuron D in an information transfer process. D will receive inputs from other neurons, let’s call them \[x_1, x_2, \dots, x_n\]. If 5 neurons passed information to D, then n would be 5. Let the weights between these n neurons and D be \[w_1, w_2, \dots, w_n\], and the amount of information they pass to D be \[x_1, x_2, \dots, x_n\]. Then the sum of the information input to D would be \[y = \sum_{i=1}^{n} w_i \cdot x_i\], which is the amount of information carried by each neuron multiplied by its weight and added together. If D has a threshold of 0, then if this value is greater than 0, D will pass information to other neurons, otherwise nothing will happen. We call this algorithm a “perceptron”, which means “cognitive machine”.
The best part about perceptrons is that they can improve their accuracy on their own, without the need to change the algorithm. The data used by a perceptron is labeled +1 or -1, depending on whether it actually conveys information. If a perceptron with a certain weight correctly predicts 50 labels out of 100 pieces of data, its accuracy is 50%. In the three-step algorithm we saw earlier in this documentation, increasing the accuracy would require fundamentally tearing the algorithm apart, but in the case of a perceptron, we don’t need to do that; we can simply tweak the weights according to a known formula and the accuracy will increase. After repeating this process a few times, no matter how many times you change the weights, the accuracy will no longer increase, at which point the learning ends and the final model is formed.
Let’s go back to the Japanese flag problem. Let’s say that for each given picture data, we feed the perceptron the width and length of the picture and the percentage of the picture that is covered by the red circle. The machine will already have a label of +1 if the given data is the Japanese flag and -1 if it is not. If the result of training with a weight of 0.4 for the width and length and 0.2 for the proportion of circles is less accurate than the given data, the machine will modify the weights, train again, and compare again. Repeat this process to get the most accurate weights, say 0.5, -0.2, and 0.7. Then, when a new picture is fed in, the machine will calculate “horizontal length * 0.5 – vertical length * -0.2 + proportion of circles * 0.7” and classify it as +1 if this value is greater than 0 and -1 otherwise.
It is self-evident that the perceptron is a black box model. While the perceptron is an interesting concept, there is no logical connection between the Japanese flag and the perceptron. Nevertheless, the reason why this neural network model has become a global phenomenon is that it is surprisingly accurate. Let’s look at a face recognition problem that is much more difficult than the Japanese flag identification problem. The face recognition technology used in Apple’s latest product, the iPhone X, is based on deep learning with perceptrons. It has a face recognition rate of 99.9% accuracy. This is not a phenomenon that can be explained by classical logic; it is a phenomenon that deserves to be called a revolution.
In this way, machine learning is very different from the AI we see in science fiction movies. Humans are not simply learning machines with problem-solving capabilities. Humans can sometimes explain phenomena without data, and solve one problem while solving a completely new one, which is clearly impossible for modern AI. Therefore, we need to treat machine learning as an efficient problem-solving process in the era of AI for the Fourth Industrial Revolution, and not be overly fearful of it.
I’m a blog writer. I want to write articles that touch people’s hearts. I love Coca-Cola, coffee, reading and traveling. I hope you find happiness through my writing.