US 11,113,627 B2Granted

Characterisation of data sets corresponding to dynamical statistical systems using machine learning

AssigneeNoble Artificial Intelligence, Inc.

InventorMatthew Chase Levy

Life of this patent

Estimating the term…

Loading patent…

Abstract

Machine learning is performed on input data representing a dynamical statistical system of entities having plural primary variables that vary time. A distribution function over time of the density of entities in a phase space, whose dimensions are the primary variables and secondary variables dependent on the rate of change of the primary variables, is derived and encoded as a sum of contour functions over time describing the contour in phase space of plural phaseons which are entities of a model of the dynamical statistical system that are localised in the phase space. Machine learning is performed on the encoded distribution function and/or at least one field in the effective configuration space whose dimensions are the primary variables, derived from the encoded distribution function. The encoding of the distribution function provides a representation which improves the performance of the machine learning techniques by simplifying hyperparameter optimisation.

Claims21

Dependent claims

A method of performing machine learning on a data set representing a dynamical statistical system of entities having plural primary variables that vary with time, the entities comprising one or more species,

the method treating the primary variables as dimensions in an effective configuration space, and treating secondary variables that are dependent on a rate of change of each of the primary variables as variables in an effective momentum space, and treating the effective configuration space and the effective momentum space together as a phase space,
the method comprising,
with respect to each of the one or more species of the entities:
providing a set of training feature vectors corresponding to training examples of phaseons in a plurality of classes;
training a machine learning architecture to perform classification based, at least in part, on the training feature vectors;
deriving a distribution function over time of a density of entities in the phase space;
encoding the distribution function as a sum of contour functions over time describing the contour in phase space of plural phaseons which are entities of a model of the dynamical statistical system that are localised in the phase space;
and
performing, using the machine learning architecture, machine learning on the encoded distribution function or at least one field in the effective configuration space derived from the encoded distribution function, with respect to each of the one or more species of the entities;
wherein:
feature vectors are derived for the plural phaseons based, at least in part, on the encoded distribution function or the at least one field in the effective configuration space;
and
the machine learning performs classification of the plural phaseons into one or more of the plurality of classes using the feature vectors.

10 more elements

A system of performing machine learning on a data set representing a dynamical statistical system of entities comprising one or more species and having plural primary variables that vary with time, the system comprising:

a computing device that includes a computer-readable storage medium storing a computer program, the computer program being configured to cause the computer apparatus to:
execute a method which treats the primary variables as dimensions in an effective configuration space, which treats secondary variables that are dependent on a rate of change of each of the primary variables as variables in an effective momentum space, and which treats the effective configuration space and the effective momentum space together as a phase space,
with respect to each of the one or more species of the entities:
providing a set of training feature vectors corresponding to examples of phaseons in a plurality of classes;
training a machine learning architecture to perform classification based, at least in part, on the training feature vectors;
derive a distribution function over time of a density of entities in the phase space;
encode the distribution function as a sum of contour functions over time describing the contour in phase space of plural phaseons which are entities of a model of the dynamical statistical system that are localised in the phase space;
and
perform, using the machine learning architecture, machine learning on the encoded distribution function or at least one field in the effective configuration space derived from the encoded distribution function, with respect to each of the species of the entities;
wherein:
feature vectors are derived for the plural phaseons based, at least in part, on the encoded distribution function or the at least one field in the effective configuration space;
and
the machine learning performs classification of the plural phaseons using the feature vectors.

10 more elements

Dependencies are inferred from the claim text.