DESIGN TOOLS
Micron technology glossary

Deep learning

Deep learning is a subset of machine learning that has become more important over the years. Deep learning models are crucial for powering artificial intelligence (AI) and advancing cutting-edge technologies. Read more to learn what deep learning truly is, how it functions and how it is being used across a range of industries.

What is deep learning?

Deep learning definition: Deep learning is a subset of machine learning that uses artificial neural networks with multiple layers to analyze and interpret complex data.

These neural networks, composed of numerous processing layers, work collaboratively to extract and process various features and patterns from datasets.

​​Neural networks are a type of computer system modeled on the human brain, designed to mimic human intelligence in the way it processes information. Deep learning uses these deep neural networks to process information in a way that replicates human cognitive functions. Through this process, computers can learn from data and improve over time, leading to more accurate outcomes.

While some neural networks are relatively simple, deep learning relies on deep, complex neural networks with many layers. These networks are used to simulate intricate decision-making processes.

How does deep learning work?

Deep learning is based on algorithms that are trained on large sets of labeled data. These algorithms learn to associate data with specific labels and can apply this categorization to new data. Once trained, a deep learning algorithm can make predictions based on the data it processes.

Here is how it works: Deep learning uses artificial neural networks to learn from data. These networks, modeled after the human brain, enable complex processing and learning from data similarly to how humans learn information.

Data is passed through layers of interconnected neural networks to produce results. In deep learning, data is processed through many layers of nodes (neurons). Each node receives input data, processes it and passes the output to the next layer of nodes. This layered approach results in sophisticated and nuanced outputs. The systems learn during processing (training) and from the outputs, recognizing patterns that can be used later (inference) and further mimicking the human brain’s learning and processing capabilities.

What is the history of deep learning?

Deep learning has a rich technological history with many stages of development, beginning in the early 1940s with computer models based on neural networks. Since then, it has undergone significant evolution.

  • 1940s, foundational concepts: The development of deep learning can be traced back to the 1940s. In 1943, Walter Pitts and Warren McCulloch created a computer model based on neural networks, using a combination of algorithms and mathematics they called “threshold logic” to mimic the human thought process. This foundational work laid the groundwork for later advancements in neural networks and deep learning.
  • 1960s, early development: Deep learning began to take shape in the 1960s when the first models were highlighted. Key developments included Kelley’s continuous backpropagation model and Dreyfus’s simpler chain rule, marking the earliest iterations of deep learning models.
  • 1970s, CNNs and the AI winter: Throughout the 1970s, interest in and funding for ​​artificial intelligence research declined, a period known as the AI winter. Despite this lull, significant advancements were made in convolutional neural networks (CNNs), symbolic logic and backpropagation, laying the groundwork for future deep learning development.
  • 1980s and 1990s, faster processing: In the 1980s and 1990s, as computers began to process data more quickly and efficiently, deep learning regained momentum. Enhanced processing capabilities allowed neural networks to learn faster, driving stronger research and progress.
  • 2000s, refining the systems: The 2000s saw further refinement of deep learning systems. Discoveries about how later layers inform earlier layers led to advancements in short-term memory and layer-by-layer training. In 2009, Fei-Fei Li launched a database of over 14 million labeled images, significantly advancing the training and development of deep learning models.
  • 2010s, breakthroughs and widespread adoption: The 2010s marked a period of significant breakthroughs and widespread adoption of deep learning. Key developments included the rise of deep neural networks, such as convolutional neural networks and recurrent neural networks, which achieved remarkable success in image and speech recognition tasks. The introduction of generative adversarial networks in 2014 by Ian Goodfellow revolutionized the field by enabling the generation of realistic images and data. Additionally, the availability of large, labeled datasets and advancements in computational power, particularly through GPUs, accelerated deep learning research and applications.
  • 2020s, advanced architectures and applications: The 2020s have seen the emergence of more advanced architectures, such as transformers, which have become the foundation for state-of-the-art models in natural language processing (NLP) and other domains. Deep learning continues to expand into various fields, including healthcare, autonomous driving and robotics, demonstrating its versatility and impact. Ethical considerations and the need for explainable AI have become increasingly important as deep learning models are integrated into critical decision-making processes.

What are key types of deep learning?

Deep learning algorithms are incredibly complex, with multiple layers of neural networks constructing the model. These layers process information sequentially, with each layer’s output informing the next. There are various ways that deep learning models can be constructed and applied, each suited to different tasks.

  • Convolutional neural networks (CNNs) are a type of deep learning model that learns directly from the data it is fed, specializing in recognition tasks. They are particularly powerful in processing image data but also effective in handling speech and audio data.
  • Recurrent neural networks (RNNs) are effective in ​natural language processing and speech recognition. They are designed to understand verbal context as well as temporal dependencies, which helps models understand how past events influence future outcomes. This capability enables more accurate predictions and better decision-making, which in turn makes RNNs suitable for tasks such as language translation and time-series forecasting.
  • Generative adversarial networks (GANs) are used in AI to generate creative outputs that resemble training data. They consist of two neural networks, the generator and the discriminator, which work together in an adversarial manner. The generator creates artificial data, while the discriminator evaluates the authenticity of the data. This dynamic leads to the generation of highly realistic images, videos and other content.
  • Deep reinforcement learning (DRL) is commonly used for robotics and game play. It enables computers to learn how to behave in an environment by using machine learning techniques to optimize their actions based on feedback from the environment. DRL models learn through trial and error, improving their performance over time by maximizing rewards and minimizing penalties.

How is deep learning used?

Semiautonomous vehicles are a prime example of deep learning in action. These vehicles use deep neural networks to process vast amounts of data from sensors, cameras and radar systems in real time. By analyzing this data, AI can recognize objects, predict the behavior of pedestrians and other vehicles, and make complex driving decisions.

High-performance memory is crucial for supporting AI workloads, enabling the rapid processing and storage of large datasets needed for training deep learning networks. This memory ensures that the system can manage the intensive computational demands, allowing the AI to learn from new data and improve its accuracy and safety over time. This technology is essential for developing self-driving cars that can navigate diverse environments and handle various driving scenarios with minimal human intervention.

Micron’s leadership products are at the forefront in providing the high-performance memory solutions that power these advanced AI applications. Innovative memory solutions, such as HBM3E and 1β (1-beta) DRAM, are designed to meet the demanding requirements of AI workloads. Micron’s HBM3E memory, for instance, offers exceptional bandwidth and capacity, significantly enhancing the performance of AI systems by enabling faster data processing and reducing power consumption. Additionally, Micron’s comprehensive portfolio, including high-capacity server RDIMMs and Gen5 PCIe® NVMe™ SSDs, ensures that AI applications can scale efficiently and effectively.

Frequently asked questions

Deep learning FAQs

Machine learning and deep learning are both subsets of AI, but they differ in their complexity, data requirements, and processing methods. Deep learning is a specialized subset of machine learning, designed to enhance and make machine learning systems more efficient. 

Deep learning uses highly interconnected neural networks that work together to process data. Unlike traditional machine learning, which often requires significant human intervention for feature extraction and training, deep learning automates these processes, making it a more autonomous form of AI technology.

Deep learning models offer several advantages over traditional machine learning models: 

Efficiency: Deep learning models can process and analyze large datasets more efficiently, reducing the need for extensive human intervention. 

Complexity: They can learn complex relationships between features in a dataset, making them powerful tools for tasks such as image and speech recognition, natural language processing and more. 

Scalability: Deep learning models are highly scalable and improve their performance as they process more data, making them ideal for big data applications. 

Autonomy: These models require minimal human intervention for training, allowing for more autonomous and adaptive AI systems.