Artificial neuron
lastmod 2018-08-17created 2015-12-239 edits
An artificial neuron is a processor with only simple computational functions, devised on the basis of neuroanatomical facts.
Structure
Cognitive scientists interested in human thought and cognition, and researchers interested in new computational models, devised processors (artificial neurons) with only simple computational functions on the basis of neuroanatomical facts. They then proposed a computational model in the form of a network that connects these processors through weighted channels (data paths). This proposed model is called an artificial neural network. The artificial neuron, the basic building block of an artificial neural network, is a processor with the structure shown in the figure. Unlike the computers we usually use, in which one or a few processors with complex functions perform all computation and control, the basic concept of an artificial neural network is that a great many artificial neurons with simple computational functions are interconnected to store and process information.
Artificial_neuron.png
Function
The function of an artificial neuron, the processor that makes up an artificial neural network, is as follows. It simply multiplies the input signal x=[x1, x2, … , xn] by the connection weights w=[w1, w2, … , wn], sums all the products, and then applies a nonlinear function f to the result. Here, for the same input x, the output takes different values depending on w. So the information can be seen as stored in the connection weight vector w. The output y is computed by the following equation.
\[ y = f \left ( \sum_{i=1}^n x_n \cdot w_n - \theta \cdot w_0 \right ) \]
Here f is a function with nonlinear characteristics1, called the activation function or transfer function. θ is the neuron's own threshold or bias, which keeps the neuron from activating when the value of Σ(xi·wi) is smaller than the threshold.
See also
Favorite site
- Wikipedia (en) article on artificial neurons
- Description of the neuron in mathematical neural networks
- [Recommended] Neural network theory : Oh Chang-seok 2 (includes explanations of several activation functions)