Artificial Intelligence (AI) models are a critical component of any AI system. They are the brains that support AI-enabled systems to analyse, decide, and perform the action for which the system is developed. In recent years, a range of AI models have been developed to support a plethora of use cases, each model being suited for specific requirements.
Traditional AI models are rule-based systems, relying on explicitly programmed instructions without ongoing learning derived by applying intelligence to data inputs received during live operations. Further advances have resulted in AI models that are categorised into three broad groups: supervised learning, unsupervised learning, and reinforcement learning. Supervised and unsupervised learning models help in making predictions, identifying patterns, and clustering data into meaningful groups, while in reinforcement learning models learn from their environment using a system of rewards and punishments for the actions that a system takes. For every correct output generated, it is rewarded and for incorrect output, it is penalised.
A neural network is an advanced type of machine learning model inspired by how a human brain works. It can process large and complex datasets comprising different data forms that include text, audio, photos, videos, and sensor feeds.
Generally, a neural network is made up of three main layers, which are input, hidden, and output layers. The input layer is the initial layer in a network that accepts input data and the nodes of this layer correspond to a feature or dimension (aspect) of the data. Hidden layers sit between the input and output layers, are responsible for complex computations, and are critical to network performance. The output layer is the final layer responsible for producing predictions.
Meanwhile, AIoT and edge AI applications, particularly those related to mission-critical scenarios, often demand real-time responses. These applications often require models with extremely low latency, minimal energy consumption, and reduced memory usage. For instance, in the automotive domain, autonomous vehicles rely on object detection and classification, which necessitate lightweight and efficient AI models. However, deploying the most sophisticated and complex AI models directly on AIoT devices can result in slower inference speed, leading to slow responses.
To overcome this, AI models undergo various opimisation techniques such as compression, quantisation (which lowers the precision of the weights and activations in a neural network by converting floating-point numbers into lower precision numbers such as eight bit integers), and pruning (which lowers the number of parameters in a neural network by removing unnecessary connections and parameters).
In this report, we explore the different kinds of model that have been developed and the applications to which they are best suited. The report discusses the most common neural networks used in edge AI and AIoT contexts. We have also discussed some of the common machine learning algorithms that are used across various industrial IoT applications.
