US 12,340,572 B2
Device and method to adapt a pretrained machine learning system to target data that has different distribution than the training data without the necessity of human annotations on target data
Chaithanya Kumar Mummadi, Pittsburgh, PA (US); Evgeny Levinkov, Stuttgart (DE); Jan Hendrik Metzen, Boeblingen (DE); Kilian Rambach, Stuttgart (DE); and Robin Hutmacher, Renningen (DE)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on May 18, 2022, as Appl. No. 17/747,361.
Claims priority of application No. 21179755 (EP), filed on Jun. 16, 2021.
Prior Publication US 2022/0406046 A1, Dec. 22, 2022
Int. Cl. G06V 10/00 (2022.01); G06V 10/764 (2022.01); G06V 10/82 (2022.01)
CPC G06V 10/82 (2022.01) [G06V 10/764 (2022.01)] 10 Claims
OG exemplary drawing
 
1. A computer implemented augmented machine learning system, comprising:
a parameterized input transformation module; and
a pretrained learning system configured to classify images;
wherein an output of the parameterized input transformation module is connected with an input of the pretrained machine learning system, and the input transformation module is configured to at least linearly transform its input, and further configured to input its transformed input to the pretrained machine learning system;
wherein the input transformation module is configured to at least partially undo a domain shift of inputs of the input transformation module such that outputs of the input transformation module come relatively close to an original input of a training data distribution on which the pretrained machine learning system has been trained;
wherein the input transformation module includes a further machine learning system which is connected in series with a linear transformation module, the further machine learning system being configured to non-linearly transform the input of the input transformation module,
wherein the linear transformation module is configured to linearly transform its input depending on parameters characterizing a linear transformation, and
wherein between the further machine learning system and the linear transformation module, an addition module is interconnected, and the addition module includes at least two inputs, and a first input of the at least two inputs is configured to receive outputs of the further machine learning system and a second input of the at least two inputs is configured to receive the input of the input transformation module, and the addition module is configured to weighted sum its inputs and output the sum to the linear transformation module.