Neuromorphic computing and neural networks might sound similar, but they function differently. Neuromorphic computing imitates the human brain with artificial neurons that fire and form connections with other neurons, which can individually break, reroute and reconnect. Neural networks also employ artificial neurons but in a net of links that activate in layers with data moving in only one direction. Both approaches are currently used to bring AI to edge computing.
How Do Neural Networks Function?
Artificial neurons in a neural network are called nodes. Nodes are arranged in layers – input, output and hidden. The input layer receives information for the network to process. The output layer signals the response to the processed information. The hidden layer performs various processing functions and sits between the input and output layers. Depending on complexity, a network may have one or any number of individual hidden layers.
Nodes are connected to each other by channels that are assigned weights, a numerical value that is multiplied to incoming data. If the resulting computation is below a set threshold value, the node doesn’t pass data to the next later. If the result is higher than the threshold, the node “fires” and passes data forward. In the output layer, the node with the highest value determines the result.
When a neural network is undergoing training, weights and threshold values are set to random. Data is fed to the input layer and the resulting output is analyzed. If the network reached an incorrect conclusion, the correct output is taught through a process called backwards propagation. In this process, data is fed to the output layer and nodes fire backward through the network, adjusting weight and threshold values to match the correct conclusion.
Training of the network continues through forward and backward data propagation until the network reaches the correct output in most cases. Depending on complexity, training a neural network can take anywhere from hours to months.
Neural Network Applications
Neural networks are used to recognize patterns. They can be trained to read handwriting, detect disease in medical images or perform facial recognition. Neural networks are used in satellites and drones for image recognition. They can been applied in power system restoration to identify where and in what order to restore power over a service area. In recent news, a neural network has been trained to differentiate between Middle Stone Age and Later Stone Age hunter-gatherer tools – a feat difficult for human researchers to replicate.
The neural network market is expected to reach $38.71 billion globally by 2023.
- Case Studies
- Quickstart Guides
- Press Releases
- White Papers