Convolutional neural networks

Convolutional neural networks (CNN) are a specialized subset of neural networks, that offer advanced structures, and ​​machine learning capabilities, enabling computers to perform tasks autonomously that once required human intervention. Join Micron as we delve into the world of convolutional neural networks and discover how they are driving innovation and transforming our world. 

What are convolutional neural networks?

Convolutional neural networks definition: CNNs are a specialized type of artificial neural network that uses ​deep learning algorithms to analyze and process data directly. 

​​Neural networks are a category of ​​artificial intelligence technology, with multiple subsets including convolutional neural networks, feedforward neural networks, and backpropagation neural networks. Though they are all based on the fundamental concept of mimicking human brain functions, they differ in their configurations, data input methods and learning processes.

Convolutional neural networks differ from other kinds of neural networks in that they learn directly from the data they receive and their structural design. Unlike traditional neural networks where each node is connected to every node in the subsequent layer, CNNs have fewer interlayer connections. This unique architecture makes CNNs particularly effective for tasks like image recognition and classification, which is why they are frequently used for ​computer vision.

How do convolutional neural networks work?

Convolutional neural networks fundamentally work the same way that other neural networks do. They have an input layer where data is introduced; several processing layers where data is analyzed; and an output layer where results, analysis or content is delivered to the user.

CNNs are also especially well-equipped to handle visual data, including images and videos because they use a technique called parameter sharing. Instead of every node being connected to every node in the next layer, elements of the data are categorized so similar data is not processed repeatedly.

But how do CNNs work differently from other types of neural networks? The layering system works slightly differently, with three main types of layers.

• The convolutional layer is always the first layer, but there can be multiple convolutional layers. It processes the input data by applying filters to detect features. The three main components of this layer are the input data, the filter, and the output feature map. The filter scans the image to create a feature map, identifying patterns and features. Key parameters for this process include: 
  • Number of filters: Determines how many feature maps are created. 
  • Stride: Defines how far the filter moves, affecting the output size. 
  • Padding: Impacts the output size compared to the input.  
• The pooling layer simplifies the data from the convolutional layer by reducing its size and the number of parameters while retaining crucial information. The data is divided into similar fields and aggregated. The two main types of pooling are: 
  • Max Pooling: Selects the pixel with the maximum value. 
  • Average Pooling: Calculates the average value of the pixels. 
• Unlike most layers in convolutional neural networks which are only partially connected, each node in the fully connected layer is connected to every node in the previous layer, allowing for the classification of features into a final output. 

What is the history of convolutional neural networks?

While development of neural network technology began as long ago as the 1940s, convolutional neural networks were not developed until much later.  

• 1980, neo cognitron: In 1980, Kunihiko Fukushima introduced the neocognitron, which had two types of layers. The S-layer, which became known as the convolutional layer, applied filters to detect features in the input data. The C-layer, similar to modern pooling layers, reduced the spatial dimensions of the data, helping to consolidate information and reduce computational complexity. 
• Late 1980s, development of CNNs: The next step in the development of CNNs from neural networks was LeNet. Yann LeCun and others developed this system as a digit recognition model, but the architecture of convolutional, pooling and fully connected layering laid the groundwork for convolutional neural networks. 
• 2000s, breakthrough technology: With two decades of development complete, CNNs became an important technology in the 2000s. Graphic processing units were accelerated, which was crucial for furthering machine learning.  
• 2010s, high-performance CNNs: In the 2010s, CNNs reached superhuman levels of image recognition while developing generative AI technologies raised the prospect of enhanced content generation based on CNNs. 
• 2020s, modern advancements: The 2020s have seen significant advancements in CNN architectures, such as the development of ConvNeXt. This new architecture combines the strengths of traditional convolutional networks with modern design elements inspired by transformers, achieving impressive performance in tasks like object detection and semantic segmentation. Additionally, the integration of CNNs with other AI technologies has expanded their applications in various fields, including medical imaging, autonomous driving and more. 

What are key types of convolutional neural networks?

Conventional neural networks are themselves a subset of neural network technology, specifically operating to process spatial data. Though there are many approaches to using and applying CNNs and many CNN models that have specific features and advantages, CNNs do not have subsets or types. 

How are convolutional neural networks used?

Convolutional neural networks power much of the AI we are coming to rely on, so this technology has wide-ranging applications. Some of the most promising are medical imaging and self-driving cars. 

CNNs offer a route to more efficient, automated identification and diagnosis of medical conditions through enhanced medical imaging. Using convolutional neural networks to process visual media like medical X-rays and CT scans enables medical professionals to harness the power of AI to highlight issues that may be missed. By offloading the work of visual assessment to highly effective CNN models, medical imaging processing allows for smoother and more efficient diagnosis. 

Another emerging technology that uses convolutional neural networks is autonomous vehicles. Self-driving cars need CNNs to power the complex interoperability on which the technology relies. Autonomous vehicles must continuously assess objects and movements in real time, often processing multiple data types simultaneously. Powerful CNNs make ongoing assessments possible.