Conditional Normalizing Flow Priors for Multi-Modal CVAE Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing conditional variational autoencoders (CVAEs) face issues such as over-regularization, difficulty in capturing multi-modal distributions, and latent variable collapse, leading to unimodal generations and poor learning of probability distributions, particularly in complex environments like real-world traffic scenes.

Innovation Solution

A machine learnable system using a variational autoencoder with flexible priors and normalizing flows, where the encoder and decoder functions generate parameters defining a probability distribution with a predetermined variance, and a sequence of invertible normalizing flow sub-functions, improves the learning of complex probability distributions, especially in latent spaces with multiple modes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a standard Gaussian prior is used in CVAE, then the model structure is simple, but the model suffers from over-regularization and difficulty in capturing multi-modal distributions

Engineering Contradiction:
Improvemodel structureVSAvoidability to capture multi-modal distributions
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent changes the prior distribution parameter from a standard Gaussian to a flexible normalizing flow-based distribution. This allows the model to adapt to complex multi-modal distributions while maintaining a structured approach through the flow-based transformation framework.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces a normalizing flow as an intermediary transformation between the latent space and the prior distribution. This mediator enables the model to handle complex distributions without directly modifying the core CVAE architecture, thus maintaining simplicity while improving adaptability.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If CVAE is used for generation, then the model can handle conditional data, but latent variable collapse occurs and minor modes are not captured

Engineering Contradiction:
Improveconditional generation capabilityVSAvoidlatent variable utilization
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent modifies the latent variable distribution parameters by using a flexible prior with normalizing flows, which prevents collapse to a single mode and enables reliable representation of multiple conditional modes including minor modes.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces dynamic flexibility to the prior distribution through normalizing flows, allowing the latent variable distribution to adapt its shape and complexity based on the conditional input, thereby preventing collapse while maintaining conditional generation capability.

Inventive Principle:
Principle #15Dynamics

3Adaptability or versatility

If the encoder uses learnable variance, then the model has more flexibility, but training becomes unstable and learning is hindered

Engineering Contradiction:
Improveencoder flexibilityVSAvoidtraining stability
Core Design Contradiction:
Adaptability or versatilityVSEase of manufacture

Solution Approach 1:

The patent changes the variance parameter from a learnable quantity to a fixed predetermined value. This simplifies the training landscape and improves stability while the flexibility is maintained through the normalizing flow transformation in the prior distribution rather than through variance learning.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3767533B1A machine learnable system with normalizing flow
Publication Date: 2026.05.20 ROBERT BOSCH GMBH
  • EP3767533B1 patent drawingFigure 1a~1b
  • EP3767533B1 patent drawingFigure 1c~1d
  • EP3767533B1 patent drawingFigure 2a~2b

AI summary

Some embodiments are directed to a machine learnable system (110). A conditional normalizing flow function maps a latent representation (z) to a base point (e=f(z,c)) in a base space (E) conditional on conditioning data (c). The conditional normalizing flow function is a machine learnable function and trained on a set of training pairs.