Chapter 656 [Neural Network Deep Learning]
An hour later, Fang Hong came to the headquarters of Quantitative Capital again.
Chen Yu's assistant came to receive him, led him to the reception room, and said, "Mr. Fang, Mr. Chen is having a meeting in the technical department. Wait a moment. I'll inform him."
Fang Hong said, "No, just take me to his conference room, and I'll listen in."
Hearing this, Chen Yu's assistant took out his mobile phone and sent him a message, and soon Chen Yu replied. The assistant turned to look at Fang Hong with a smile and said, "Mr. Fang, this way please."
After a while, Fang Hong came to the conference room where Chen Yu was. There were more than 30 people present. Seeing a strange young man walking in, everyone looked at him curiously.
They found that Fang Hong was about the same age as their boss, Chen Yu, but the difference was that they felt an aura of superiority from Fang Hong that they didn't have at this age, which made everyone realize that this strange young man was not an ordinary person.
At this moment, Chen Yu saw Fang Hong looking at him and nodding his head. The latter smiled and quietly found a seat in the meeting room to sit and listen.
Chen Yu withdrew his gaze, looked around at the participants and continued: "...for the basic realization of artificial intelligence, the process of machine learning, simply put, is how computers learn by themselves."
"Because all calculations of computers are based on mathematical operations, any idea of machine learning, in the final analysis, is to transform a practical problem into a mathematical problem. In order for the computer to predict or recognize something, it is necessary to construct a mathematical function first. This mathematical function is called the prediction function."
It may be difficult for ordinary people to imagine that Quantitative Capital, as a diversified financial company, is even a non-bank financial investment company in the eyes of most shareholders. The head of the company is also engaged in investment transactions, but they talk about these things in the company.
But Fang Hong is very calm. This is actually quite normal. Wall Street has gathered a group of top mathematicians and physicists.
At this moment, Chen Yu turned to look at the meeting screen and said: "For example, the function of predicting a full meal can be described as [full meal = N bowls of rice], is this prediction calculation accurate? How many bowls of rice a person eats and What is the relationship between being full? Is it enough to eat one bowl or three bowls?"
"This needs to be tried in practice. If the prediction is that two bowls of rice are full, but in reality it takes three bowls of rice to be full, the error of one of the bowls is the loss. The function that describes this loss is [3-N=1], This is the loss function."
"Machine learning is the process of constantly trying to minimize this error. The method of finding the minimum loss is usually gradient descent. Once we find the minimum error, we will find that when [N=3] the error is the smallest, that is, the machine When you learn to find the real law, you will successfully solve the problem.”
Chen Yu looked at the crowd again and said: "So, machine learning is to find the laws of data. Most of the time, its essence is to project the data into the coordinate system, and then use the computer to draw a line mathematically to distinguish or simulate these data. the process of."
"Different machine learning methods are using different mathematical models to project data and draw lines. From the last century to the present, different schools have found different methods and are good at solving different problems. There are a few that have had a huge impact. Species: Linear and Logistic Regression, K-Nearest Neighbors, Decision Trees, Support Vector Machines, Bayesian Classification, and Perceptrons, etc.”
Fang Hong sat on the sidelines and listened silently. He is also half of the industry in the field of computer science, and he has the advantage of past life memory and foresight. There is no pressure to listen at this moment.
Chen Yu and the others are obviously following the neural network genre, but they have also taken a step forward and entered into intensive deep learning, and the predecessor of the neural network is the perceptron.
These three nouns are essentially playing the same thing.
But at this moment, Chen Yu said slowly: "The most basic idea of deep learning is to simulate the activity of brain neurons to construct prediction functions and loss functions. Since it is called a neural network, it must have a certain relationship with human brain neurons. The algorithmic mechanism of a single perceptron is actually simulating the operating mechanism of neurons in the brain."
A diagram of the structure of neurons in the brain appears on the screen.
"This is a neuron. We all know its structure. This is the dendrite, and this is the axon. Signals from other neurons enter the neuron through the dendrite, and then emit through the axon. This is a nerve." Yuan's operating mechanism."
"Now we mutate the tree of neurons into an input value and the axon into an output value, so the neuron becomes a graph like this. It's even simpler to convert it into a mathematical formula, [ X1+X2+X3=Y], that’s the formula.”
"That's right, it's that simple. The most complex things are often created by the simplest things. Simple 0s and 1s shape a huge computer world, and four nucleotides vacate complex life phenomena. One Simple neuronal reflexes shape our brains."
Chen Yu paused for a while, and looked around again: "The key to the problem is not how simple the basic structure is, but how we use this basic structure to build a huge world. The reason why neurons are magical is that they have an activation mechanism, the so-called threshold."
"Each dendrite of a neuron continuously receives input signals, but not every input signal can cause the axon to output signals, and each dendrite has different weights when inputting."
"For example, when you pursue a girl, you take various actions tirelessly. You send her a bouquet of flowers today, treat her to a big meal tomorrow, but you find that none of these actions can impress her. Until one day you went shopping with her for a day. , she was suddenly moved and agreed to be your girlfriend, what does that mean?"
"It shows that not all input weights are the same. Shopping with girls may have the greatest weight. Secondly, the accumulation of effects is not a linear and gradual process, but quantitative changes lead to qualitative changes."
"All inputs have no effect before a certain point, but once they reach a certain value, they are suddenly excited. Therefore, to imitate the activation characteristics of neurons, then modify the formula just now."
"Each input needs a certain weight. Add a coefficient [W] to adjust the weight in front, and add a constant in the back to adjust the threshold better, so this function becomes like this."
Fang Hong also looked at the big screen of the conference, it was a new mathematical formula.
【W1X1+W2X2+W3X3+b=Y】
Chen Yu looked at the formula on the screen and said: "In order to realize the activation process, further process the output value and add an activation function, for example, when X\u003e1, output 1; when X\u003c1, output 0, So it became like this.”
"However, this function does not look round enough and is not derivable everywhere, so it is not easy to deal with. It is replaced by the Sigmoid function. Such a simple function can handle classification problems."
"A single perceptron actually draws a line to separate two different things. A single perceptron can solve linear problems, but it can't do anything for linear inseparable problems. That means that even the simplest XOR problem is impossible. Can't handle it."
Everyone present, including Fang Hong, understood the XOR problem, and it was one of the basic calculations of computers.
At this time, Chen Yu asked himself rhetorically: "If the XOR problem cannot be resolved, isn't that the rhythm of the death penalty?"
Chen Yu immediately replied: "It's very simple, directly use the kernel function to increase the dimension. The reason why the perceptron can become the current deep learning is because it has changed from one layer to multiple layers. The depth of deep learning refers to the depth of the perceptron. There are many layers, and we usually call a neural network with more than three hidden layers a deep neural network, how does the perceptron solve the XOR problem by adding layers?"
Chen Yu looked back at the screen and took the next slide and said: "Computers have four basic logics of operation, and, or, negate, and exclusive or. I don't need to talk about this. If we put the exclusive or in a coordinate system To show that's it."
"The origin position X is 0, Y is 0, so take 0; when X=1, Y=0, if the two are different, take 1, Tongli, here is also 1, and this position X, Y are both equal to 1, so take 0, If we need to separate 0s and 1s on this graph, a straight line can’t do it.”
"What to do? It depends on the nature of the XOR operation. Mathematically speaking, the XOR operation is actually a compound operation, which can be obtained through other operations. The proof process is too complicated and I won't expand here."
"If we can use the perceptron to complete the calculation in the brackets first, and then input the result into another perceptron for the outer calculation, we can complete the puzzle operation, and then the XOR problem is so magical Solved, while solving the problem, it also solved the problem of linear inseparability."
"What does this mean? It means that no matter how complex the data is, we can fit a suitable curve to separate them by adding layers, and adding layers is the nesting of functions. In theory, no matter how complicated the problem is, we can It is combined by simple linear functions, so, theoretically speaking, multi-layer perceptrons can become a general method that can solve various machine learning problems across domains."
...
