Variational Autoencoder Training Using Sigma Points for Low-Variance Sampling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Training variational autoencoders (VAEs) for anomaly detection and sampling in high-dimensional spaces often results in high-variance gradient computations, leading to inefficient reconstruction and sampling capabilities.

Innovation Solution

The method employs sigma points, specifically mean-centered symmetric points, to determine a latent distribution in VAEs, reducing gradient variance and improving reconstruction abilities by using these points for training, rather than relying on the reparameterization trick.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If the reparameterization trick is used to simplify approximate posterior sampling in VAEs, then the gradient computation becomes more straightforward, but the sampling points may be very far from the mean in high-dimensional spaces, leading to high-variance policy gradient

Engineering Contradiction:
Improvegradient computation simplicityVSAvoidgradient variance
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The patent changes the sampling parameters by using sigma points with specific weighting schemes instead of standard random sampling. The sigma points are positioned at specific distances from the mean (e.g., sqrt(n) times the standard deviation) and assigned appropriate weights to reduce gradient variance while maintaining computational tractability.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces sigma points as an intermediary between the mean and random samples. These sigma points serve as structured intermediate representations that capture the distribution characteristics more effectively than simple random samples, thereby reducing gradient variance without sacrificing computational simplicity.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Device complexity

If single or few random samples are used in VAE setting, then the computation is simple, but the samples can be very far from the mean, especially in high dimensional spaces, leading to poor reconstruction ability

Engineering Contradiction:
Improvesampling complexityVSAvoidreconstruction precision
Core Design Contradiction:
Device complexityVSManufacturing precision

Solution Approach 1:

The patent modifies the sampling parameters by using multiple sigma points positioned at specific distances from the mean with appropriate weighting. This changes the distribution of sampling points to be more concentrated around the mean while still capturing the essential variability, thereby improving reconstruction precision without significantly increasing computational complexity.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent uses a moderate number of sigma points (more than single sampling but fewer than exhaustive sampling) to achieve sufficient reconstruction quality. This partial action approach provides a good trade-off between computational complexity and reconstruction precision, avoiding both oversampling and undersampling extremes.

Inventive Principle:
Principle #16Partial or excessive action

3Manufacturing precision

If sigma points are used instead of reparameterization trick, then the reconstruction ability of VAE is improved and latent distribution is better modeled, but the gradient computation requires high-variance policy gradient

Engineering Contradiction:
Improvereconstruction precisionVSAvoidgradient variance
Core Design Contradiction:
Manufacturing precisionVSReliability

Solution Approach 1:

The patent resolves this contradiction by carefully selecting the parameters of sigma points, specifically their distance from the mean and their weighting coefficients. By optimizing these parameters, the patent achieves both improved reconstruction precision through better latent distribution modeling and reduced gradient variance through appropriate weighting schemes.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent incorporates feedback mechanisms in the training process where the gradient variance is monitored and the sigma point parameters are adjusted accordingly. This feedback loop allows the system to adapt the sigma point configuration to minimize gradient variance while maintaining reconstruction quality.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4343626A1Device and method for training a variational autoencoder
Publication Date: 2024.03.27 ROBERT BOSCH GMBH
  • EP4343626A1 patent drawingFigure 1
  • EP4343626A1 patent drawingFigure 2
  • EP4343626A1 patent drawingFigure 3

AI summary

Computer-implemented method for training a machine learning system (60), wherein the machine learning system (60) is configured for anomaly detection and/or sampling a trajectory for a traffic participant and/or sampling of sensor signals and/or for determining a value characterizing a likelihood of an input signal with respect to a training dataset, wherein training comprises: • Determining, by an encoder (61) of the machine learning system (60) and based on a training input signal (xi), a first intermediate representation (r1) characterizing a mean of a latent distribution of a latent space (l) and a second intermediate representation (r2) characterizing a variance and/or covariance of the latent distribution; • Determining, based on the first intermediate representation (r1) and the second intermediate representation (r2), a plurality of sigma points (σ) with respect to the latent distribution; • Determining an output signal (yi), wherein the output signal (yi) is determined by providing a randomly sampled sigma point (σ) of the plurality of sigma points (σ) to a decoder (62) of the machine learning system (60); • Adapting the machine learning system (60) based on a loss value, wherein the loss value characterizes a difference between the training input signal (xi) and the output signal (yi).