US 12,468,937 B2
Device and method for training a classifier using an invertible factorization model
Volker Fischer, Renningen (DE); Chaithanya Kumar Mummadi, Heimsheim (DE); and Thomas Pfeil, Kornwestheim (DE)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on Sep. 20, 2021, as Appl. No. 17/448,110.
Claims priority of application No. 20198965 (EP), filed on Sep. 29, 2020.
Prior Publication US 2022/0101128 A1, Mar. 31, 2022
Int. Cl. G06N 3/08 (2023.01); G06F 17/18 (2006.01)
CPC G06N 3/08 (2013.01) [G06F 17/18 (2013.01)] 13 Claims
OG exemplary drawing
 
1. A computer-implemented method for training a classifier, wherein the classifier is configured to determine an output signal characterizing a classification of an input signal, the method comprising the following steps:
determining a first training input signal from a training data set, wherein the training data set comprises, for each input signal, a corresponding output signal;
determining a first latent representation including a plurality of factors based on the first training input signal using an invertible factorization model,
wherein the invertible factorization model includes:
a plurality of functions, wherein each function from the plurality of functions is continuous and continuously differentiable in at least a portion thereof, wherein the function is further configured to accept either the training input signal or at least one of the factors provided by another function of the plurality of functions as an input and wherein each function of the plurality of functions is further configured to provide at least one of the factors, wherein the at least one of the factors is either provided as at least part of the latent representation or is provided as at least part of an input of another function of the plurality of functions, wherein the plurality of factors of the latent representation are provided at least by two different functions of the plurality of functions, wherein a respective function of the plurality of functions of the invertible factorization model is invertible, wherein for each respective function there exists a corresponding inverse function, wherein each inverse function is continuous, is continuously differentiable in at least a portion thereof and is configured to determine an input of the respective function based on the at least one factor provided from the function;
determining a second latent representation by changing at least one of the factors of the first latent representation;
determining a second training input signal based on the second latent representation using the invertible factorization model using backward propagation;
training the classifier based on the second training input signal by determining with the classifier based on a second input signal an output signal and comparing the determined output signal to a respective corresponding output signal; and
providing an actuator configured to control, based on the output signal of the trained classifier, a physical action of a vehicle, a robot, or a machine.