Supervised and semi-supervised learning: Data, that has already been categorized by humans, is analyzed. Through this data, an algorithm is created, that attempts to classify new input data. The model is then trained to further improve classification, commonly through backpropagation.
Unsupervised learning: Similar to supervised learning, though no human categorization is required. The system simply classifies data according to similarities.
Reinforcement learning: Given a certain state, the system can perform multiple actions, each of which either provides positive or negative feedback (in this state the feedback is unknown though). It chooses one action and receives feedback. This feedback is taken into consideration, when choosing the next action. Through this feedback loop, the system can evaluate, which actions tend to produce the highest amount of positive feedback.
Convolutional Neural Networks: CNNs are used for recognizing patterns in a given input, such as recognizing wether or not the image contains a cat. Multiple layers extract features from a given input. The first layers extract fairly abstract features (such as edges), whereas the next layers may extract shapes and finally, further layers may extract features such as faces. At the end, one single neuron outputs the probability of the input containing the object the network was trained for. Usage: Image or sound classification
Recurrent Neural Networks: Neurons in an RNN not only get input from the previous layer, but also from themselves in the previous passes. The neurons in this type of network develop over time, so the sequence in which you feed it information influences the output. Usage: natural language, translation
Generative Adversarial Network: GANs actually consist of two networks: one generating network and one adversarial network. The generating network generates new data, based on initial input data. The adversarial network then tries to distinguish between the initial input data and the generated data. The generating network then optimizes its process, to generate data that is harder for the adversarial network to distinguish, whereas the adversarial network improves its ability to distinguish real from generated data. Usage: generation of artificial imagery, predictions about the next frames of a video
Bayesian networks are a different approach to machine learning, not related to neural networks. They are based on "Bayes' Theorem", which describes the probability of an event based on already established knowledge of related conditions. Bayesian networks aim to model conditional dependence, and therefore causation. They can be used to achieve a certain sense of abstraction through use of causal conclusion. Compared to ANNs they do not require as much input data, but causal connections must be designed, as they cannot work "blindly". Usage: medical diagnosis, decision making
Sources: