Argmax Image Generation Model for Stable Likelihood Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The limited availability of training data poses a challenge for achieving high accuracy in machine learning-based image processing, particularly in applications like autonomous driving, where the model needs to be reliable and robust, and collecting real-world data can be expensive or dangerous.

Innovation Solution

An image generation model using an argmax transformation to discretize continuous latent feature vectors, allowing for efficient training and accurate generation of synthetic images that represent the training dataset without adding uncorrelated noise, and a stochastic inverse transformation for determining conformance values.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If noise is added to the generative process in VAE training, then the model can be trained more stably, but the generated images become less representative of the training dataset and quality deteriorates

Engineering Contradiction:
Improvetraining stabilityVSAvoidimage generation quality
Core Design Contradiction:
ReliabilityVSManufacturing precision

Solution Approach 1:

The patent changes the parameterization approach by using the log-sum-exp trick to reformulate the likelihood calculation, allowing stable training without adding uncorrelated noise to the generative process. This mathematical transformation enables gradient computation while maintaining exact likelihood evaluation, resolving the contradiction between training stability and image quality.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If real-world training data is collected to improve model accuracy, then model reliability improves, but the cost and danger increase significantly

Engineering Contradiction:
Improvemodel accuracyVSAvoiddata collection cost
Core Design Contradiction:
ReliabilityVSEase of manufacture

Solution Approach 1:

The patent uses the trained image generation model to synthesize additional training data that copies the statistical properties of real-world data. This synthetic data can be used to augment or replace expensive real-world collected data, maintaining model accuracy while reducing data collection costs and dangers.

Inventive Principle:
Principle #26Copying

3Productivity

If the image generation model uses continuous transformations, then training efficiency improves, but the generated discrete image data loses precision

Engineering Contradiction:
Improvetraining efficiencyVSAvoiddiscrete image data accuracy
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent introduces the log-sum-exp transformation as an intermediary mathematical operation that bridges continuous latent representations and discrete image data. This intermediary enables exact likelihood computation and gradient flow through the discretization process, maintaining both training efficiency and discrete data precision.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3975038B1An image generation model based on log-likelihood
Publication Date: 2026.03.11 ROBERT BOSCH GMBH
  • EP3975038B1 patent drawingFigure 1~2
  • EP3975038B1 patent drawingFigure 3
  • EP3975038B1 patent drawingFigure 4a

AI summary

The invention relates to a computer-implemented method (700) of training an image generation model. The image generation model comprises an argmax transformation configured to compute a discrete index feature indicating an index of a feature of the continuous feature vector with an extreme value. The image generation model is trained using a log-likelihood optimization. This involves obtaining a value of the index feature for the training image, sampling values of the continuous feature vector given the value of the index feature according to a stochastic inverse transformation of the argmax transformation, and determining a likelihood contribution of the argmax transformation for the log-likelihood based on a probability that the stochastic inverse transformation generates the values of the continuous feature vector given the value of the index feature.