What is AI?

When most people today think of AI, they will often imagine text and image generators. But even the simplest of algorithms is technically an example of AI in the literal sense.

The broadness of AI and its multiple use cases have led to several subdomains, including machine learning and deep learning, as shown in Figure 1.

Figure 1 The relationship between different AI subdomains.

The majority of AI used for embedded applications is machine learning, the subdomain where machines and algorithms “learn” how to solve a problem from data; for example, a vehicle recognizing a pedestrian vs. an obstacle by analyzing image data for common patterns. A machine learning model learns from training data, which may be labeled with ground truth information (i.e. verified, accurate data) to better identify which patterns to learn from. This training process enables machine learning models to discern patterns in the data, which they can use to make future inferences.

Within the field of machine learning, deep learning has become one of its most popular implementations given its ability to solve highly complex problems accurately, although doing so requires plenty of computing resources. Deep learning uses multilayer neural networks, which are data models inspired by the neurons in the human brain. Neural networks enable developers to solve problems where the patterns are too complex to discern or write custom rules for.