Глоссариум по искусственному интеллекту: 2500 терминов. Том 2 - страница 38

Discriminative model is a model that predicts labels from a set of one or more features. More formally, discriminative models define the conditional probability of an output given the features and weights; that is (output|features, weights). For example, a model that predicts whether an email is spam from features and weights is a discriminative model. The vast majority of supervised learning models, including classification and regression models, are discriminative models. Contrast with generative model413.

Discriminator is a system that determines whether examples are real or fake. The subsystem within a generative adversarial network that determines whether the examples created by the generator are real or fake414.

Disparate impact – making decisions about people that impact different population subgroups disproportionately. This usually refers to situations where an algorithmic decision-making process harms or benefits some subgroups more than others415.

Disparate treatment – factoring subjects’ sensitive attributes into an algorithmic decision-making process such that different subgroups of people are treated differently416.

Dissemination of information – actions aimed at obtaining information by an indefinite circle of persons or transferring information to an indefinite circle of persons417.

Dissemination of personal data – actions aimed at disclosing personal data to an indefinite circle of persons418.

Distributed artificial intelligence (DAI) (also decentralized artificial intelligence) is a subfield of artificial intelligence research dedicated to the development of distributed solutions for problems. DAI is closely related to and a predecessor of the field of multi-agent systems419.

Distributed registry technologies (Blockchain) are algorithms and protocols for decentralized storage and processing of transactions structured as a sequence of linked blocks without the possibility of their subsequent change420.

Distribution series are series of absolute and relative numbers that characterize the distribution of population units according to a qualitative (attributive) or quantitative attribute. Distribution series built on a quantitative basis are called variational421.

Divisive clustering – see hierarchical clustering422,423.

Documentation generically, any information on the structure, contents, and layout of a data file. Sometimes called «technical documentation» or «a codebook». Documentation may be considered a specialized form of metadata424.

Documented information – information recorded on a material carrier by means of documentation with details that make it possible to determine such information, or, in cases established by the legislation of the Russian Federation, its material carrier425.

Downsampling – overloaded term that can mean either of the following: Reducing the amount of information in a feature in order to train a model more efficiently. For example, before training an image recognition model, downsampling high-resolution images to a lower-resolution format. Training on a disproportionately low percentage of over-represented class examples in order to improve model training on under-represented classes. For example, in a class-imbalanced dataset, models tend to learn a lot about the majority class and not enough about the minority class. Downsampling helps balance the amount of training on the majority and minority classes