Issue link: https://resources.mouser.com/i/1481902
STMicroelectronics 2022 7 Figure 3: The three types of Machine Learning. (Source: Mouser Electronics) Categories of ML The goal of ML is a value-generating predictive model. Here are the three main categories of ML. • Image and Voice recognition: Machine recognition adds a new dimension to human interaction with devices and makes them more autonomous—more able to perceive and react to their environment without an explicit prompt. • Adaptive solutions: Nowadays, ML provides continuous self-learning about new data to incrementally improve performance and efficiency. • Complex monitoring: When a single event ML model does not fully account for the behavior of a complex application, a multi-step ML model can be implemented for connected sub-tasks with intermediate inference goals. Types of ML There is no "one-size-fits-all" approach to ML because every application is unique. Different types of problems favor different algorithms and training strategies. For example, a people counting system that uses a camera has very different hardware and ML algorithm requirements than a predictive maintenance system that fuses accelerometer and electrical sensor data. ST's STEVAL-STWINKT1B is an example of a powerful hardware kit that integrates an MCU and various sensors including sensors with embedded machine learning. As an ML algorithm learns on training data, its predictive error decreases. Once high enough accuracy is achieved on the training data, the model is deployed within the product's embedded program. Algorithm selection and model development are key ML success factors, as well as the quality and quantity of training data. Shown above are the three types of machine learning (Figure 3). Figure 8 Supervised Learning • Makes machine Learn explicitly • Data with clearly defined output is given • Direct feedback is given • Predicts outcome/future • Resolves classification and regression problems • Machines understands the data (Identifies patterns/ structures) • Evolution is qualitative or indirect • Does not predict/find anything specific • An approach to AI • Reward based learning • Learning from +ve & +ve reinforcement • Machine Learns how to act in a certain environment • To maximize rewards Unsupervised Learning Inputs Inputs Inputs Outputs Training Rewards Outputs Outputs Reinforcement Learning Figure 8 Supervised Learning • Makes machine Learn explicitly • Data with clearly defined output is given • Direct feedback is given • Predicts outcome/future • Resolves classification and regression problems • Machines understands the data (Identifies patterns/ structures) • Evolution is qualitative or indirect • Does not predict/find anything specific • An approach to AI • Reward based learning • Learning from +ve & +ve reinforcement • Machine Learns how to act in a certain environment • To maximize rewards Unsupervised Learning Inputs Inputs Inputs Outputs Training Rewards Outputs Outputs Reinforcement Learning Figure 8 Supervised Learning • Makes machine Learn explicitly • Data with clearly defined output is given • Direct feedback is given • Predicts outcome/future • Resolves classification and regression problems • Machines understands the data (Identifies patterns/ structures) • Evolution is qualitative or indirect • Does not predict/find anything specific • An approach to AI • Reward based learning • Learning from +ve & +ve reinforcement • Machine Learns how to act in a certain environment • To maximize rewards Unsupervised Learning Inputs Inputs Inputs Outputs Training Rewards Outputs Outputs Reinforcement Learning