3. Machine learning

Machine learning is a procedure to learn from examples and, more specifically, it is a field of optimizing system parameters, which are defined on an architecture, to meet the evaluation criteria using a set of training examples. We often use statistical techniques to give computers the ability to "learn." Once the intended goal of learning is met, we may use the resulting system to automatically predict the category of unseen data, to estimate location in the feature space, or to generate artificial examples depending on different applications. Machine learning algorithms are typically classified into three broad categories: supervised learning, unsupervised learning, and reinforcement learning.

For supervised learning problems, the training data comprises examples of the input vectors along with their corresponding target vectors. When the target vectors are categorical, the problems are known as classification or pattern recognition, and when the target vectors are real-valued, the problems are known as regression. Loss or distance functions are defined between the current output vector and the target vector for each input vector, and optimization is performed to minimize the loss over all training examples. By teaching the system with known input and target pairs, we expect to respond correctly even if unseen data are presented to the trained system.

For unsupervised learning problems, no targets are defined so that the training data consist of only a set of input vectors. The goal of unsupervised learning is to automatically discover "interesting statistical structure" in the data. It can also be explained as latent knowledge discovery from examples, and a variety of clustering algorithms are canonical examples of unsupervised learning.

Reinforcement learning [3] is to learn how to act or behave in a given situation for given reward or penalty signals. In this type of learning, a state for current status is defined and environment, usually a criterion function, evaluates the current state to generate a proper reward or penalty action through a set of policies. Instead of having exact target values, it learns with critics.

Deep learning [4] has been inspired from human brain and has been proving its powerful ability in detection, classification, segmentation, key point estimation, to activity classification. It generally consists of huge number of parameters with multiple nonlinear layers. Deep learning architectures include two popular categories: convolutional neural networks (CNN) for automatic feature extraction and recurrent neural networks (RNN) for sequence estimation. They have been applied to computer vision, speech recognition, natural language processing, audio recognition, social network filtering, machine translation, and bioinformatics with outstanding performances. In addition, generative models such as variational encoders and generative adversarial networks (GAN) are also becoming popular with their artificial sample generation capability.
