US 12,271,796 B2
Machine learning system to generate output using first and second latent spaces
Christos Louizos, Utrecht (NL); Welling Max, Bussum (NL); and Xiahan Shi, Renningen (DE)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on Jun. 4, 2020, as Appl. No. 16/893,083.
Claims priority of application No. 19180246 (EP), filed on Jun. 14, 2019.
Prior Publication US 2020/0394506 A1, Dec. 17, 2020
Int. Cl. G06F 9/54 (2006.01); G06F 18/214 (2023.01); G06N 3/063 (2023.01); G06N 3/08 (2023.01); G06N 20/10 (2019.01)
CPC G06N 20/10 (2019.01) [G06F 9/545 (2013.01); G06F 18/214 (2023.01); G06N 3/063 (2013.01); G06N 3/08 (2013.01)] 17 Claims
OG exemplary drawing
 
1. A machine learning system for an autonomous device, the machine learning system configured to classify objects represented in input data instances, the machine learning system comprising:
an input interface configured to obtain the input data instances, the input data instances including sensor data from a sensor of the autonomous device, the sensor detecting an environment of the autonomous device, the autonomous device including at least one of: a vehicle, or a robot;
a storage element configured to store a set of reference data instances, the set of reference data instances including reference sensor data; and
a processor circuit configured to perform a method of classifying the objects represented in the input data instances, including:
mapping the input data instances to first latent input vectors in a first latent space according to a first function,
determining similarities between the first latent input vectors and first latent reference vectors, the first latent reference vectors obtained according to the first function from the set of reference data instances;
for a first input data instance of the input data instances:
identifying a number of reference data instances in the set of reference data instances as parents of the first input data instance based upon at least one of the determined similarities,
determining a second latent input vector in a second latent space for the first input data instance from second latent reference vectors, the second latent reference vectors obtained according to a second function from the number of reference data instances identified as parents,
determining a first output from the second latent input vector by applying a third function to the second latent input vector, the output including a classification of an object of the objects, represented in the sensor data, in the environment of the autonomous device, the classification classifying the object as at least one of: a second vehicle, a person, or a cyclist,
wherein the third function and at least one of the first function and the second function are machine learnable functions; and
generating a control signal to automatically modify a movement of the autonomous device as a function of the classification of the object, where the modifying of the movement includes at least one of: stopping the autonomous device, decelerating the autonomous device, or steering the autonomous device; and
for a second input data instance of the plurality of input data instances:
identifying that no parents of the second input data instance exist as reference data instances in the set of reference data instances based on the determined similarities;
determining a default second latent input vector in the second latent space; and
determining a default output based on the default second latent input vector.