Hypothesis Language

A Hypothesis Language is the language that can describes the hypotheses that are represent as the output set used by a machine learning system.

• AKA: Representation Language.
• Example(s):
  • …
• Counter-Example(s):
  • Observation Language (input).
• See: First-Order Logic; Hypothesis Space; Inductive Logic Programming; Propositional Language; Instance-Based Learning; Clustering; Neural Network; Decision Tree; Rule Set; Graphical Model.

---

---

References

2011

• (Blockeel, 2017) ⇒ Hendrik Blockeel. (2017). "Hypothesis Language" In: (Sammut & Webb, 2017)
  • QUOTE: The hypothesis language used by a machine learning system is the language in which the hypotheses (also referred to as patterns or models) it outputs are described.

    (...) Most machine learning algorithms can be seen as a procedure for deriving one or more hypotheses from a set of observations. Both the input (the observations) and the output (the hypotheses) need to be described in some particular language. This language is respectively called the Observation Language or the hypothesis language. These terms are mostly used in the context of symbolic learning, where these languages are often more complex than in subsymbolic or statistical learning. For instance, hypothesis languages have received a lot of attention in the field of Inductive Logic Programming, where systems typically take as one of their input parameters a declarative specification of the hypothesis language they are supposed to use (which is typically a strict subset of full clausal logic). Such a specification is also called a Language Bias.

    (...) The hypothesis language used obviously depends on the learning task that is performed. For instance, in predictive learning, the output is typically a function, and thus the hypothesis language must be able to represent functions; whereas in clustering the language must have constructs for representing clusters (sets of points). Even for one and the same goal, different languages may be used; for instance, decision trees and rule sets can typically represent the same type of functions, so the difference between these two is mostly syntactic.

---