DESIGN TOOLS
Micron technology glossary

Neural networks

Neural networks, a fundamental aspect of artificial intelligence (AI) , enable computers to process information like humans. Artificial intelligence emulates the workings of a human, and neural networks replicate the cognitive processes and behaviors of humans.

With the prominence and integration of artificial intelligence growing in daily life, neural networks have become an increasingly important technology.

What is a neural network?

Neural network definition: A neural network is a computer system modeled on the human brain and designed to mimic the human brain’s structure and function. It processes information and makes decisions like human intelligence does.

Neural networks consist of interconnected nodes, or neurons, arranged in layers. These layers enable the network to process data in a way that resembles how the human brain does so. Neural networks continuously learn and adapt by improving their accuracy and output through experience and data.

Neural networks adapt and learn as they go. As ever-evolving systems, neural networks are powerful tools used across various industries and applications, and they are  capable of learning from data, decisions, and obstacles they encounter.

How do neural networks work?

Artificial neural networks consist of layers of nodes arranged in complex networks and designed to process data and produce outputs. These outputs are typically predictive and analytical, derived from the input data through understanding and categorization.

The simplest neural networks are composed of three interconnected layers:

  • Input layer: This layer introduces information. The input nodes process, analyze and categorize the data before passing it along to the next layer.
  • Hidden layer: This layer receives information from the input layer or other preceding hidden layers. Here, the output from the previous layer is further analyzed, processed and passed on.
  • Output layer: This is the final stage, where the processed data is returned as the result.

This process may seem simple, but complex neural networks can consist of several hidden layers with millions of neurons interlinked. Deep neural networks can be extremely complex and typically need more training than other machine learning models.

What is the history of neural networks?

Artificial intelligence is a relatively new technology, but the history of neural networks is quite long. Indeed, the evolution of technological thought on neural networks and of neural network technologies has been taking place for longer than we might expect. Here are some key milestones:

  • 1943, origins: Warren McCulloch and Walter Pitts explored how neural networks could be applied to computing, laying the groundwork for future development.
  • Late 1940s, expansion of the theory: Neural networks began to be considered in machine learning. D. O. Hebb proposed a mechanism of neural plasticity, known as Hebbian learning, which later informed the development of neural networks in AI. Hebbian learning is often summarized by the phrase, “Cells that fire together, wire together.”
  • 1958, beginning of computer learning: Frank Rosenblatt developed the perceptron, marking the advent of artificia l neural networks. This was a significant milestone, leading to increased investment in deep learning research.
  • 1967, multilayer perceptron: Shun’ici Amari published the first multilayer perceptron trained by stochastic gradient descent, an optimized function method.
  • 1980s, convolutional neural networks: Kunihiko Fukushima introduced the neocognitron, an early form of convolutional neural network (CNN) architecture that became fundamental for computer vision. The development of backpropagation in 1986 enabled the training of deep neural networks, overcoming the limitations of the perceptron and revitalizing interest in neural networks. This innovation paved the way for significant advancements, including Yann LeCun’s seminal 1989 paper on training algorithms for neural network architectures, which marked a major step forward in the successful implementation of neural networks in AI.
  • 1992, support vector machines and ensemble methods: The 1990s saw the introduction of support vector machines (SVMs) by Vladimir Vapnik and colleagues, and these machines became a popular method for classification tasks. Additionally, Leo Breiman developed the random forest algorithm in 1995, improving model accuracy through ensemble learning.
  • 2000s, deep learning and CNNs: The 2000s witnessed the widespread adoption of convolutional neural networks in various applications, including image classification and object detection. CNNs’ ability to capture spatial hierarchies and recognize patterns revolutionized the field of computer vision. In 2006, Geoffrey Hinton and Ruslan Salakhutdinov published a paper on deep belief networks, probabilistic models that can learn complex dependencies. This paper renewed interest in deep learning.
  • 2010s, transformers and LLMs: Introduction of the transformer model in 2017 revolutionized natural language processing by enabling models to understand grammatical dependencies in language. This architecture became the foundation for large language models (LLMs) like BERT and GPT, significantly advancing the field. OpenAI’s GPT-3, released in 2020, demonstrated the power of LLMs with 175 billion parameters and showcased their ability to generate humanlike text.
  • 2020s, advanced models: The 2020s have seen continued advancements in generative AI , with models like DALL-E and Stable Diffusion enabling the creation of realistic images and text. These developments have brought AI into the wider public consciousness and highlighted its commercial potential. In a significant leap forward, Google introduced an AI agent capable of producing highly accurate 15-day weather forecasts. This new tool, developed by DeepMind, highlights the power of AI in practical applications and offers superior predictions that can help in planning and mitigating damage from extreme weather events.

What are key types of neural networks?

Neural networks consist of layers of interconnected nodes, or neurons. The interactions between these and the flow of data through them differentiate the various types of neural networks.

Different types of neural networks work in slightly different ways to reach predictive and analytical outputs:

  • Convolutional neural networks(CNNs) are a type of neural network that learns directly from data. They are typically used for recognizing items, categories and patterns. With this recognition capability, CNNs are useful for classifying data into categories, series and lists and for making predictions.
  • The primary distinction of feedforward neural networks is the unidirectional flow of data. These networks do not form loops; instead, they pass information forward to other nodes and layers.
  • Unlike feedforward neural networks, which pass information forward only, backpropagation neural networks use a learning process to analyze errors and adjust outcomes. This method improves the network ability to make accurate predictions by learning complex patterns. Backpropagation neural networks use corrective feedback loops to analyze and enhance their own output.

How are neural networks used?

Neural networks are revolutionizing various fields by enhancing accuracy and efficiency. In healthcare, they analyze medical images to detect diseases, while in self-driving cars, they process sensor data to enable safe navigation and decision-making.

  • Medical imaging: Neural networks are extensively used in healthcare for analyzing medical images such as X-rays, MRIs and CT scans. They can detect anomalies and diagnose diseases with high accuracy, often faster and more reliably than human professionals.
  • Self-driving cars: Neural networks power the decision-making processes in self-driving cars. They help the vehicle understand its surroundings by processing data from sensors and cameras, enabling it to navigate, avoid obstacles and make real-time driving decisions.

Frequently asked questions

Neural networks FAQs

Neural networks are trained using large datasets. During training, these networks are fed vast amounts of information, which they process to learn patterns and relationships within the data.

A key method used in training neural networks is backpropagation, a continuous feedback technique that adjusts parameters within the neural network during training based on the error between the output and the expected result. By iteratively refining these parameters, neural networks improve their accuracy, efficiency and predictive capabilities. As they encounter more data, they adapt and become better at recognizing complex patterns, making them powerful tools in various applications.

Deep learning is a subset of machine learning that uses multilayered neural networks. These networks are constructed in complex layers to facilitate deep learning and to simulate the complex decision-making process of human intelligence.

Neural networks take a variety of forms, so here are a few important examples of neural networks:

  • Feedforward neural networks: These networks pass information in one direction, from input to output without processing loops.
  • Backpropagation neural networks: These networks learn from their own mistakes using a system of reprocessing loops to mitigate errors and inaccuracies.
  • Recurrent neural networks: These networks can remember previous outputs when receiving new inputs, making them useful for sequential data.
  • Perceptron: Introduced by Frank Rosenblatt, a perceptron is a type of artificial neural network designed to detect patterns in data.  It consists of a single layer of neurons, where each neuron acts as a binary classifier.
  • Modular neural networks: These networks consist of different modules that operate independently to complete separate tasks.