When it comes to artificial intelligence, or AI for short, we are confronted with a plethora of different terms. We are talking about weak and strong AI, training, inferencing, machine learning, deep learning, neural networks and much more.
In the first step, we go back to the beginning, to the fundamental question: How do you get a model or construct of 0 and 1 to “think” for itself? What possibilities are opened up by training a model in its application?
Training Models
Model training is a process in which an AI system runs through large amounts of data and “learns” to adapt and fine-tune its internal parameters. The goal is to create a model that can recognize patterns and relationships in the data in order to make predictions or decisions. This process can be very computationally intensive and often requires specialized hardware such as GPUs (graphics cards), future NPUs (neural processing units) or FPGAs (programmable processors). Simpler models can also be trained with conventional, high-performance CPUs, which has a positive effect on power consumption and investment costs.
Inferencing – Applying Models in Practice
After a model has been trained, the process of “inferencing” follows: the trained model processes new, unknown data in order to draw predictions or conclusions from the data. Inferencing can therefore be seen as the “application mode” of an AI model, in which it uses the knowledge it has learned to perform practical tasks such as identifying objects in images or understanding voice commands. The graphics card or CPU of the “edge” device (end device) usually serves as the hardware. Current devices are also often equipped with the aforementioned NPUs as part of the CPU. Additional AI accelerators such as Nvidia Jetson or Intel’s Gaudi 3 are another option.
Types of AI Models
AI models can be classified in different ways, depending on various criteria such as the learning algorithm, the application, the structure or the technology used. We will limit ourselves to three basic categories of AI models: machine learning, deep learning and neural networks as a form of deep learning. There are also a large number of sub-models or hybrids, such as the “generative transformative AI” known from ChatGPT as a sub-form of deep learning. As a rule, the model that best suits the use case is used.
Machine Learning
Machine learning is the classic form of AI models and has been in use for some time. Training is carried out with “labeled” training data in order to recognize patterns in the data and make predictions from them. Examples include linear regression or decision trees. A distinction is also made between supervised learning, unsupervised learning and reinforcement learning. One example of supervised learning is quality control as a form of deep learning in the manufacturing industry. Here, image recognition is used to analyze the quality of products in real time, defects or deviations are detected and defective products are sorted out if necessary.
Another common industrial application is “predictive maintenance” or “proactive” maintenance. Sensors installed on and in the machines are used to continuously monitor and analyze the current condition of the machines. The underlying algorithms learn from this data to recognize patterns that indicate impending defects or failures in order to prevent them. Maintenance or servicing can therefore be carried out in good time before a defect occurs. AI thus helps to save costs and extend the service life of machines, in some cases significantly.
Deep Learning and Neural Networks
Deep learning is a newer form and further development of machine learning that aims to independently recognize patterns in large amounts of data, similar to what the human brain does. Networks of algorithms called “neural networks” are often used because they are modeled on the structure and function of the human brain. These neural networks consist of layers of nodes or neurons. The difference to classic machine learning models is that these layers or nodes are connected to each other and the model continues to learn independently by running through the layers. Each layer takes the output values from the previous layer and calculates new values in order to pass these values on to the next layer. At the end of this process chain, the network can make decisions or predictions based on what it has learned.
Use Case: Autonomous Driving
The best-known example from the automotive industry at the moment is certainly autonomous driving. Deep learning is used there, for example, to monitor the vehicle’s surroundings using cameras and sensors. The collected data is analyzed live in real time by neural networks (see also “Inferencing” above) in order to recognize other vehicles, pedestrians, traffic signs and, if necessary, road conditions. This information helps the vehicle to navigate more safely and make decisions such as whether to stop or swerve.
Use Case: Industrial Robots
Another example is improving the capabilities of industrial robots. With the help of deep learning models, they learn to perform complex and precise tasks such as gripping unknown objects, welding or assembling parts of variable shape and size. Other AI techniques such as reinforcement learning can be used to train robots in simulated environments before they are used in real production environments. These advanced learning methods allow robots to adapt more easily to different production conditions and changing environments, increasing the flexibility and efficiency of automation technology.
Summary
In summary, it can be said that artificial intelligence has the potential to positively change industry in many areas. Be it in the area of more efficient and cost-effective production, product quality assurance or better logistics. Of course, it also makes it possible to develop new products such as self-driving cars or intelligent robots.
In the area of edge IoT devices in particular, manufacturers such as Giada, DT Research and others are developing new products based on the new Intel CPU generations, which are equipped with NPUs in addition to high performance from the 14th generation onwards.
Selected devices such as the DT340T rugged tablet or the LT355 rugged notebook from DT Research can also be equipped with Nvidia RTX graphics cards to further increase AI performance. Giada will also soon be launching a device based on Nvidia Jetson. Further devices will follow in the near future.