Glossary of Machine Intelligence Terms

A term interchangeable with “MI.” AI was the original term often used to classify the technology now often dubbed MI. Some organizations prefer the term MI because it is less likely to evoke ties to science fiction. 

MI systems that can perceive, learn from, abstract, and act on data in images and video to perform visual tasks (e.g., navigating an environment, labeling a photo). Computer vision systems typically consist of enabling technologies and techniques—sensors, cameras, and image processors—and machine learning algorithms. Applications of computer vision include object identification (i.e., in photos and videos), image and video tagging, biometrics for security and other purposes, sentiment analysis, facial recognition, and scene analysis.

An early form of MI in which researchers codified human experts’ knowledge as sets of logic-based rules and programmed these rules into computers, enabling them to perform straightforward tasks with precision and scale. These systems lost popularity during the 1980s due to their rigidity and poor handling of changing or ambiguous circumstances. 

A field concerned with producing machines able to autonomously perform tasks that would normally require human intelligence by giving them the ability to perceive, learn from, abstract, and act using data. 

A technique to achieve MI in which computers are given the ability to learn without being explic- itly programmed using algorithms that can parse, learn from, and make determinations based on data. 

MI systems that can perceive, learn from, abstract, and act on data in text and audio in natural language (i.e., conversational speech that does not adhere to rigid rules of syntax, grammar, etc.) to perform language-related tasks (e.g., generate text, translate audio, etc.). Applications include machine translation, speech recognition, language detection, sentiment analysis, document analysis, virtual assistants, report generation, news article writing, and insight summarization.

A recently developed form of MI that enables machines to perform simple, rules-based, high-volume tasks—often clerical or “back office.” Software “bots” observe clerical staff performing these tasks and develop rules for repeating them, enabling the execution of simple tasks at scale and with precision. 

An approach to training machine learning algorithms in which the algorithms are given input variables and output (or “tagged” variables) from which to learn patterns and make inferences. Supervised machine learning is useful for classification (assigning items to categories) and regression (finding values, such as price or weight) problems. 

An approach to training machine learning algorithms in which the algorithms are given only input variables from which they uncover patterns and make inferences on their own. Unsupervised machine learning is good for clustering problems (grouping people, behaviors, types, etc.) and classification problems (finding rules in data). 

Data which have not been given a label to make them useful for training super- vised machine learning algorithms. The vast majority of data in the world are unlabeled, making them relatively less expensive to use or obtain than tagged or labeled data.