.png)
By Ionică Alexandra Maria
What is an artificial neural network?
An artificial neural network is a type of computational architecture that makes decisions similarly to the human brain, consisting of a series of processing units called "nodes" that form the mathematical model of the neuron. This uses processes that mimic the path of information in biological neurons, with the nodes transmitting data to each other as neurons transmit electrical impulses.
Each network consists of layers of artificial neurons - an input layer, one or more hidden layers, and an output layer. Each node is connected to others and has a certain weight associated with it. If the output of a node is above the previously specified threshold value, the node transmits data to the next layer of the network. Otherwise, it is deactivated and stops providing data.
Neural networks are based on training data, which is used to improve their accuracy over time. Thus, they are a powerful tool in computer science and artificial intelligence, allowing data to be categorized and clustered at high speed, making language models such as ChatGPT possible.
History of artificial neural networks
The concept of neural networks can be traced back to a mathematical paper dating back to 1943, when Warren S. McCulloch and Walter Pitts published "A logical calculus of the ideas immanent in nervous activity", research that aimed to understand how the human brain can produce complex patterns through neurons or brain cells.
A few years later, in 1949, Donald Hebb wrote the paper "The Organization of Behavior", where he outlined the idea that neural networks get stronger each time they are used, a key concept in how humans accumulate information.
The Perceptron, which resulted from Frank Rosenblatt's research in 1958, was built in hardware, used to classify a set of information, and is the oldest neural network still in use.
In 1959, Bernard Widrow and Marcian Hoff developed the models called ADALINE and MADALINE, which utilize several ADAptive LINear Elements. ADALINE was designed to recognize binary patterns, while MADALINE, which is still in use, utilizes an adaptive filter to eliminate echoes on telephone lines.
After that, the attention given to this concept waned, and it has been the target of researchers again since the 80s. Artificial Intelligence research has been conducted using this type of network since the 1990s.
How do neural networks work?
The artificial neurons in an Artificial Neural Network are called units and are arranged in a series of layers that concentrate the entire Artificial Neural Network into a system. A layer can consist of anywhere from a dozen to millions of units, which varies depending on their need to learn hidden patterns in the dataset. The input layer receives data from the outside that the neural network must either analyze or remember. It then traverses one or more layers that transform the input data into valuable information for the output layer. Finally, the last layer provides an output in the form of the network's response to the input data provided.
Thus, in the case of biological neurons, the cell body processes the impulses, the dendrites carry the information to the nucleus, and the axon transfers it to other neurons. When it comes to artificial neural networks, input nodes receive the first signals, hidden layer nodes compute the weighted sum, and output nodes compute the final output, using the results of the hidden layer and the activation function. In the course of transferring data from one unit to another, the network learns more and more about the data, eventually facilitating the output. Connections can be optimized to an extent by weight allocation and are tuned during the process by adjusting these weights to improve performance.
Biological Neuron | Artificial neuron |
Dendrite | Inputs |
Cell nucleus or Soma | Nodes |
Synapses | Weights |
Axon | Output |
Difference between ANN, CNN, and RNN
An artificial neural network (ANN) is a group of multiple perceptrons or neurons present in each layer that processes inputs only in one direction through input nodes, and is one of the simplest versions of neural networks.
Advantages: Information storage across the entire network; ability to work with incomplete knowledge; fault tolerance; distributed memory.
Disadvantages: Hardware dependency; inexplicable network behavior.
Convolutional neural networks (CNN) utilize the variation of perceptrons located on multiple layers, containing one or more convolutional layers, which create feature maps that record a portion of an image to be divided into rectangles and sent for nonlinear processing.
Advantages: High accuracy in image recognition problems; automatic detection of important features without human intervention.
Disadvantages: It does not encode object position and orientation; need for lots of training data.
Recurrent Neural Networks (RNN) have a more complex structure, transmitting information in multiple directions. Each node in the model acts as a memory cell, continuing to compute and implement operations.
Advantages: Remembers each piece of information over time; extends the effective neighborhood of pixels.
Disadvantages: Creating an RNN is a difficult task; cannot process very long sequences in some cases.
References
Artificial Neural Networks and its Applications - GeeksforGeeks
What is a neural network? | Types of neural networks | Cloudflare
Difference between ANN, CNN and RNN - GeeksforGeeks
留言