US 12,269,177 B2
Machine learning based on a probability distribution of sensor data
Priyank Jaini, Amsterdam (NL); Lars Holdijk, Brummen (NL); and Max Welling, Bussum (NL)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on May 16, 2022, as Appl. No. 17/745,290.
Claims priority of application No. 21178046 (EP), filed on Jun. 7, 2021.
Prior Publication US 2022/0388172 A1, Dec. 8, 2022
Int. Cl. G06N 20/00 (2019.01); B25J 9/16 (2006.01)
CPC B25J 9/1697 (2013.01) [B25J 9/163 (2013.01); G06N 20/00 (2019.01)] 15 Claims
OG exemplary drawing
 
1. A computer-implemented method of training a machine learnable model for controlling or monitoring a computer-controlled system, the machine learnable model being configured to make inferences based on a probability distribution of sensor data, the sensor data representing measurements of: (i) one or more physical quantities of the computer-controlled system, or (ii) an environment of the computer-controlled system, and the machine learnable model being configured to account for one or more symmetries in the probability distribution of the sensor data imposed by: (i) the computer-controlled system, or (ii) the environment of the computer-controlled system, the method comprising:
sampling multiple samples of the sensor data according to the probability distribution by:
sampling initial values for the multiple samples from a source probability distribution, wherein the source probability distribution is invariant to the one or more symmetries,
iteratively evolving the multiple samples, including evolving each selected sample based on similarities of the selected sample to the multiple samples, wherein the similarities are computed according to a kernel function, wherein the kernel function is equivariant to the one or more symmetries, and wherein the selected sample is evolved by computing an attraction term and a repulsion term, and wherein:
the attraction term is computed as a weighted sum of gradient directions of the probability distribution for the multiple samples, wherein the gradient directions are weighed according to the similarities, and the probability distribution is configured to be invariant to the one or more symmetries;
the repulsion term is computed as a sum of respective gradient directions of the kernel function for the multiple samples given the selected sample; and
updating model parameters of the machine learnable model based on the evolved multiple samples.