Learning method, corresponding system, device and computer program product to update classifier model parameters of a classification device
AssigneeSTMicroelectronics S.r.l.
InventorsEmanuele Plebani, Danilo Pietro Pau
Life of this patent
Estimating the term…
Abstract
A classification device receives sensor data from a set of sensors and generates, using a context classifier having a set of classifier model parameters, a set of raw predictions based on the received sensor data. Temporal filtering and heuristic filtering are applied to the raw predictions, producing filtered predictions. A prediction error is generated from the filtered predictions, and model parameters of the set of classifier model parameters are updated based on said prediction error. The classification device may be a wearable device.
Claims28
A method, comprising:
- receiving sensor data from a set of sensors including at least one inertial data sensor;
- generating, using a context classifier of a neural network having a set of classifier model parameters of a user-activity classification model, a set of raw predictions based on the received sensor data;
- and
- generating one or more control signals to control a user interface based on the set of raw predictions, wherein the generating the set of raw predictions includes:
- applying to the raw predictions temporal filtering and heuristic filtering, producing filtered predictions;
- generating a prediction error from the filtered predictions;
- and
- updating model parameters of the set of classifier model parameters based on the prediction error;
- wherein
- the heuristic filtering includes selecting transitions between a source class and a destination class in the context classifier for which a number of predictions of the destination class reaching a confirmation threshold have been detected over a temporal window,
- the temporal window has a defined duration, and
- the confirmation threshold is a threshold number of predictions.
A system, comprising:
- a memory;
- and
- circuitry coupled to the memory, the circuitry including neural-network classification circuitry and control circuity, wherein the classification circuitry, in operation:
- receives inertial sensor data;
- generates, using a set of classifier model parameters of a user-activity classification model, a set of raw predictions based on the received inertial sensor data;
- filters the raw predictions using temporal filtering and heuristic filtering, producing filtered predictions;
- generates a prediction error from said filtered predictions;
- and
- updates model parameters of said set of classifier model parameters based on said prediction error;
- and
- the control circuitry, in operation, generates one or more control signals to control a user interface based on the set of raw predictions;
- wherein
- the heuristic filtering includes selecting transitions between a source class and a destination class in the context classifier for which a number of predictions of the destination class reaching a confirmation threshold have been detected over a temporal window,
- the temporal window has a defined duration, and
- the confirmation threshold is a threshold number of predictions.
A non-transitory computer-readable medium having contents which configure processing circuitry including neural network classification circuitry to perform a method, the method comprising:
- generating, using a user-activity classification model including a set of classifier model parameters, a set of raw predictions based on inertial sensor data;
- and
- generating one or more control signals to control a user interface based on the set of raw predictions, wherein the generating the set of raw predictions includes:
- applying to said raw predictions temporal filtering and heuristic filtering, producing filtered predictions;
- generating a prediction error from said filtered predictions;
- and
- updating model parameters of said set of classifier model parameters based on said prediction error;
- wherein
- said heuristic filtering includes selecting transitions between a source class and a destination class for which a number of predictions of the destination class reaching a confirmation threshold have been detected over a temporal window,
- the temporal window has a defined duration, and
- the confirmation threshold is a threshold number of predictions.
Dependencies are inferred from the claim text.