Machine Learning: Overview
Why Learning? Learning is a core of property of being intelligent. Hence Machine learning is a core subarea of Artificial Intelligence. There is a need for programs to behave intelligently: Adapt to new situations, environment Make decisions on new data Autonomously improve behaviour Predict what will happen next
What is Learning? It s all about improving Gaining new knowledge Acquiring new skills Reinforcing existing knowledge But not always about new things Recognising mistakes Finding and forgetting wrong beliefs
How do we learn? Knowledge acquisition Learn by experience, e.g., memorising telephone numbers. Skill refinement Learn by practicing, e.g., sports, music. Taking advice Learn by provided knowledge, e.g., from a teacher. Problem Solving Use experience of the problem solution and apply it to similar problem in the future. Induction Try to understand general rules from examples. Can require a teacher to provide some classification of the examples. E.g., social norms. Discovery Gain new knowledge without the aid of a teacher. Analogy Transferring skills. One knows how to handle one situation and applies this knowledge in a different situation.
Goal of Machine Learning Machine learning studies computer algorithms for learning to do things like complete a task make accurate predictions behave intelligently Machine learning is generally about performing better in the future Basis are previous experiences given examples already collected data external instructions Abstractly Machine Learning can be seen as trying to learn a function that returns a correct or an optimal result in the future.
Some Example Applications Handwriting recognition categorize images of handwritten characters by the letters represented Face dedection find faces in images Customer segmentation predict what a customer might be interested in Fraud detection identify credit card transactions which may be fraudulent in nature Weather forecast predict whether the sun will shine tomorrow Search result optimisation order and correct results of a search
of Machine Learning The following three are the main algorithm types for machine learning. Supervised Learning Un-supervised Learning Reinforcement Learning In addition (depending on author) there are other additional or intermediate types (e.g., semi-supervised learning).
Supervised Learning We always have full observations of causes and effects (input and output). Trying to predict a function from labelled data. Example: Optical Character Recognition Using a ground truth set of character images and their meaning. Algorithms: Decision Trees, Support Vector Machines, Case-based Reasoning,...
Un-supervised Learning We have unlabelled data. Try to detect interesting relations in data. Example: Find a face in an image. We have no examples of faces in images. Cluster pixels e.g. by proximity of colour. Algorithms: Data clustering, Association rule learning,...
Reinforcement Learning Try to optimise a function by its expected return value. Specify what to do, but not how to do it Example: Robot moving in unknown terrain. Is punished when it pumps into something. Agent will learn the difference between a Algorithms: Temporal Difference, Q-Learning, Monte Carlo Method,...
Symbolic vs Sub-Symbolic Question is whether or not we should be able to retrieve the function that has been learned. A long-standing argument in AI between symbolic vs. sub-symbolic Do we just want to create an artefact that can solve problems? Or do we also want to learn how the problem is solved by observing the artefact? Example algorithms: Symbolic: Decision Trees, Support Vector Machines. Sub-symbolic: Art. Neural Nets, Genetic Algorithms.