AI Popular Science 18
Essential Algorithms in Machine Learning
Neural Networks
The 21st century is an information age, as well as a big data age. For traditional methods and tools, processing and analyzing massive amounts of data is a huge challenge. At this time, neural networks emerged as a powerful machine learning tool, demonstrating great potential in solving big data problems. So, how did neural networks come about, what are their specific concepts and structures, and how should they be trained? In this issue, we will start from the background of the birth of neural networks and explore the concepts and structures of neural networks in depth.
neuralnetwork
1. Background of Neural Networks
Figure 1 Basic Diagram of Neuron Structure
We all know that the brain is composed of neurons, and its working principle is to convert external stimuli into electrical signals that are then transmitted to neurons, which judge whether they reach the activation threshold, thus outputting “excitatory” or “inhibitory” electrical signals, and finally synthesizing various signals in the neural center to respond to external stimuli.
Based on the understanding of human neurons, scientists invented the “artificial neuron mathematical model”: its principle is to multiply each input by different weight values (w0, w1, w2) to represent the degree of influence of different inputs on the final output; after weighted summation, a bias b is added to activate the threshold for neuron transmission; finally, a nonlinear transformation is completed through an activation function to output the final value.
Figure 2 Mathematical Model of Artificial Neuron
By comparing the relationship between biological neurons and artificial neurons, as shown in the following figure, we can roughly see how artificial neurons mimic biological neurons in their work.
Figure 3 Artificial Neuron Structure and Function Diagram
shenjingwangluo
2. Concept of Neural Networks
From the above, we can see that artificial neural networks simulate the structure and function of the brain’s nervous system using computational models, establishing a neural network system with a large number of simple computational units (artificial neurons).
shenjingwangluo
3. Structure of Neural Networks
(The term “neural network” below refers to “artificial neural networks”)
(1) Single-Layer Perceptron
A single-layer perceptron consists of an input layer and an output layer with a specific activation function (sigmoid). The weights (w) in the perceptron are obtained through training and can be regarded as a type of logistic regression model, capable of performing linear classification tasks.
Figure 4 Neuron Model
Figure 5 Linear Classification Effect of Single-Layer Perceptron
(2) Multi-Layer Perceptron
When we add at least one hidden layer to the single-layer perceptron, which only has an input layer and an output layer, it becomes a multi-layer perceptron. By increasing the number of hidden layers and the number of neurons, it solves the XOR problem that a single-layer perceptron cannot solve and has very good nonlinear classification effects.
[Textbook Express]
Submitted by | Resource Service Department
Edited by | Library Promotion Team
Reviewed by | Head of Library Promotion Team
