Artificial intelligence (AI) is a simulation of intelligent human behavior. It is the software in a computer or system designed to perceive its environment, make decisions, and take action. Building a successful AI system involves understanding the entire workflow and focusing on more than just training an AI model.
How Does AI Work?
The term artificial intelligence is broad and continues to expand in relevancy as more engineers incorporate AI into a wide range of applications. Consider some of the more recent applications that range from self-driving cars to robotics to natural language processing. The way these applications implement artificial intelligence will vary, but the underlying technology—deep learning or machine learning models used to create a system that can make decisions—remains constant.
The choice between deep learning or machine learning algorithms for your artificial intelligence application depends on your system’s goals and requirements.
Traditional machine learning enables the training of various classifiers, such as support vector machines (SVMs) and decision trees. Machine learning also allows for optimizing feature extraction. You can combine different approaches to determine the best arrangement for the data.
Deep learning is a specialized form of machine learning, which automates the extraction of relevant features from data. Deep learning networks often have greater predictive power than classical machine learning models, and their accuracy improves as the size of your training data increases.
Try a variety of machine learning algorithms to find the best fit for your model.
Whether you choose machine learning or deep learning, you need the ability to try a variety of algorithms and decide what works best for your application and requirements.
Keep Exploring This Topic
Essential AI Concepts
A solid artificial intelligence workflow involves understanding your data, creating a model, and designing and testing the final system on which the model will run. The following sections describe important AI concepts to consider when incorporating artificial intelligence into your work.
Data-Driven AI
At the center of most artificial intelligence applications is data. Taking raw data and making it useful for an accurate and meaningful model will most likely represent significant time spent on your AI effort. Data preparation requires domain expertise to understand the data’s critical features, which ones are unimportant, and what rare events to consider.
Data preparation and getting to labeled data is often tedious and time-consuming. The process could include augmenting data sets with synthetic data and more samples, but engineers should consider getting to clean, labeled data faster by automating the time spent labeling.
AI Modeling
Two key factors for building a successful AI model:
- Choose a set of algorithms: Are you looking at machine learning or deep learning? Starting with a complete set of algorithms and pre-built models means you are already ahead of the game by taking advantage of the broader work in the artificial intelligence community and not starting from scratch.
- Iterate on your model: This is where you identify the optimal set of parameters that will get you to the most robust and accurate model. Creating an accurate model takes time. Fortunately, adding more hardware— such as running on one or multiple GPUs—can significantly speed up the time to train models with all combinations of parameters, input data, and layers.
Simulation in AI Design
AI models typically exist within large, complex systems. For example, in automated driving systems, artificial intelligence for perception must integrate with algorithms for localization and path planning and controls for braking, acceleration, and other components. These pieces work together to create a complete system. Complex, AI-driven systems like these require integration and simulation.
See AI in action: Detecting Oversteering in BMW Automobiles with Machine Learning
Artificial intelligence is used in different parts of a complex, robotics system.
Simulation is how it all comes together. It verifies that the pieces work together correctly. Simulation ensures the results and reactions are what you expect in every situation. It also lets you validate your model works correctly before deploying to hardware.
Real World Artificial Intelligence Applications
Since many applications use artificial intelligence, there are various deployment requirements, whether it is an ECU in a vehicle, an edge system in a chemical plant, or a cloud-based streaming system receiving data from several locations. Artificial intelligence can reside in any part of these systems, so your models need to be able to deploy and operate on any possible platform.
- Acquire data: Quickly interface with different data acquisition hardware, organize large amounts of data, or generate synthetic data when limited training data is available.
- Preprocess and label data: Build better data sets faster with preprocessing and labeling apps. Use low-code apps and built-in functions in MATLAB® to improve data quality and automatically label the ground truth.
- Build AI model: Try a variety of models and compare machine learning and deep learning algorithms to find the right solution for your application. Access hundreds of pretrained models, including models from TensorFlow™ and PyTorch®, and perform transfer learning to save time and resources.
- Visualize decisions: Gain trust in AI decisions by using explainability techniques and verifying the robustness of AI models. Techniques like LIME, Shapley, and Grad-CAM are accessible directly in MATLAB, so you do not have to write the functions manually.
- Simulate: Integrate AI models into Simulink to build artificial intelligence functionality directly into your complex systems. This integration allows engineers to simulate artificial intelligence within the entire system before deploying the model to production.
- Deploy to the edge: Identify and eliminate coding errors by automatically generating code and targeting your device. MATLAB automatically generates code for your specific target hardware so that you can integrate models into enterprise systems, clusters, and clouds or embedded hardware.
Engineers and scientists are the domain experts who provide insights that are critical to the success of AI projects. MATLAB empowers engineers and scientists to use artificial intelligence in their domain and enables collaboration across teams and organizations in various industries.
Keep Exploring This Topic
Artificial intelligence is software in a computer or system designed to perceive its environment, make decisions, and take action.
Machine learning is statistical methods that enable machines to “learn” tasks through data without explicit programming. Deep learning is a subset of machine learning using neural networks that automate feature extraction from data and often has greater predictive power that improves with larger training data sets.
The choice depends on your system’s goals, requirements, the size of data you’re processing, and the type of problem you want to solve. Typically, smaller data sets work well for machine learning algorithms, but if you have a large data set and significant processing power, try deep learning algorithms.
A solid AI workflow involves understanding your data, creating a model, and designing and testing the final system on which the model will run, including data preparation, AI modeling, simulation, and deployment.
Data preparation requires domain expertise to understand critical features and rare events, and getting to clean, labeled data faster is essential since it represents significant time in any AI effort.
Simulation verifies that all components of complex systems work together correctly, ensures results match expectations in every situation, and validates models before deploying to hardware.
Depending on the size of the model, AI can be deployed and operated on many platforms, including ECUs in vehicles, edge systems in chemical plants, or cloud-based streaming systems receiving data from multiple locations.
MATLAB enables engineers to acquire and preprocess data, build and compare AI models, visualize decisions with explainability techniques, simulate AI within complex systems, and deploy to edge devices through automatic code generation.
