Multiscale Encoder Training for Perturbation-Robust Data Compression

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current data-compression systems are prone to perturbations at irrelevant scales, leading to sensitivity issues and decreased performance in downstream tasks, as they fail to optimize for 'useful' sparsity that is both data and task relevant, and robust against smaller scale perturbations.

Innovation Solution

A computer-implemented method for training a data-compression system using an encoder neural network and memory module, where training data items are processed to generate representations and views, with loss functions based on mutual-information types to iteratively update parameters, focusing on removing data blocks determined by a cutoff scale, thereby increasing the scale in each iteration and using sliding windows to select views.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If data-compression systems use traditional compression methods, then compression is achieved, but the system becomes sensitive to perturbations at irrelevant scales

Engineering Contradiction:
Improverobustness against perturbationsVSAvoidsensitivity to perturbations at irrelevant scales
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent segments the frequency spectrum into multiple scales using wavelet transforms, organizing compression across different resolution levels. This segmentation allows the system to treat perturbations at different scales independently, applying appropriate compression strategies to each scale while preserving important features and filtering out noise at irrelevant scales.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different compression strategies to different scales based on their local characteristics. Important scales (those containing relevant information) receive lighter compression or preservation, while less important scales (containing noise or irrelevant information) receive stronger compression. This local quality approach makes the system robust by protecting critical information while compressing non-critical components.

Inventive Principle:
Principle #3Local quality

2Reliability

If sparsity is optimized using mathematical objectives, then robustness to perturbations increases, but performance on downstream tasks decreases

Engineering Contradiction:
Improverobustness to perturbationsVSAvoidperformance on downstream tasks
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent changes the parameterization of the compression system by using wavelet coefficients as the basis for compression rather than traditional transform coefficients. By adjusting the sparsity parameter λ and the scale parameter σ across different wavelet levels, the system can optimize both robustness and downstream task performance simultaneously, as different scales capture different aspects of the data relevant to different tasks.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces a new dimension of scale organization through wavelet transforms, moving from single-scale compression to multi-scale compression. This dimensional change allows the system to achieve robustness through sparsity in the wavelet domain while maintaining task performance by preserving information across multiple scales, effectively decoupling the trade-off between robustness and productivity.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Productivity

If all scales are compressed uniformly, then compression efficiency is achieved, but useful information at critical scales is lost

Engineering Contradiction:
Improvecompression efficiencyVSAvoidloss of useful information at critical scales
Core Design Contradiction:
ProductivityVSLoss of information

Solution Approach 1:

The patent applies non-uniform compression across different scales by adjusting the sparsity parameter λ and scale parameter σ for each wavelet level. Critical scales that contain important information are compressed less aggressively, while non-critical scales are compressed more aggressively. This local quality approach maintains compression efficiency overall while preserving useful information at critical scales.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent introduces dynamic adaptation of compression parameters across scales, where the sparsity and scale parameters are adjusted based on the importance of each scale. This dynamic approach allows the system to optimize compression efficiency while adaptively preserving information at scales that are critical for downstream tasks, rather than applying a static uniform compression strategy.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4303774A1Multiscale deep-learning method for training data-compression systems
Publication Date: 2024.01.10 DOMOHEALTH SA
  • EP4303774A1 patent drawingFigure 1
  • EP4303774A1 patent drawingFigure 2
  • EP4303774A1 patent drawingFigure 3

AI summary

The present invention concerns a computer-implemented method for training a data-compression system comprising an encoder neural network and a memory module, comprising the steps of a. receiving at least one training data item of a given scale, the training data item being laid out sequentially along one or more dimensions; b. processing the at least one training data item through the encoder neural network to generate a representation of the training data item; c. determining and storing in the memory module a plurality of views of the at least one training data item; d. processing each view of the at least one training data item through the encoder neural network to generate a view representation of the view; e. for each pair consisting of the representation and a view representation, determining a loss function depending on a mutual-information type of function; f. iterating steps a. to e. above; g. aggregating the loss functions computed at each iteration of step e. into a total loss function; h. updating a set of parameters of the encoder neural network to minimize the total loss function; characterized in that said plurality of views of the at least one training data item is generated by removing data blocks determined by a cutoff scale from the training data item; wherein in each iteration of steps a. to e. the training data item comprises the representation generated during the preceding iteration; the scale of the training data item is greater than the scale of the training data item of the preceding iteration; and, the cutoff scale used to determine the plurality of views of the at least one training data item is greater than the cutoff scale of the preceding iteration.