Deep Neural Network Compression via Filter Clustering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep neural networks (DNNs) require significant computational power and storage, making them challenging for mobile devices and vehicles, and existing pruning methods often compromise generalizability or are inefficient.

Innovation Solution

A method involving filter similarity pruning using subspace clustering and dimension reduction to compress DNNs, preserving structural integrity and reducing storage without sacrificing accuracy, through iterative processes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If deep neural networks are trained to achieve high accuracy on large datasets, then model accuracy is improved, but model size and computational resources required increase

Engineering Contradiction:
Improvemodel accuracyVSAvoidmodel size
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the neural network into a teacher network and a student network, where the teacher network performs the heavy computational task of training on large datasets while the student network stores compressed representations. This segmentation allows high accuracy to be achieved without requiring the deployed model to be large, as the student network only needs to store compressed features rather than all training data.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent extracts essential information from the training data and represents it in a compressed form within the student network. By taking out only the necessary features and representations while discarding redundant information, the model achieves high accuracy with reduced model size and computational requirements.

Inventive Principle:
Principle #2Taking out (Extraction)

2Measurement precision

If neural networks are trained on large datasets, then model accuracy improves, but training time and computational resources increase

Engineering Contradiction:
Improvemodel accuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary action by pre-training the teacher network on large datasets before deploying the student network. The heavy computational work of training is completed in advance, and the student network only needs to load and use the pre-computed compressed representations, significantly reducing the training time required for deployment.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent creates a compressed copy of the essential learning patterns from the training data and stores them in the student network. This copying approach allows the model to capture the essential information from large datasets without requiring to process all data during deployment, reducing computational time and resources.

Inventive Principle:
Principle #26Copying

3Device complexity

If model parameters are reduced for deployment, then computational resources decrease, but model accuracy may deteriorate

Engineering Contradiction:
Improvemodel sizeVSAvoidmodel accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent changes the parameter representation from storing all training data and model weights to storing compressed feature representations. This parameter change allows the model to maintain high accuracy with fewer parameters, as the compressed representations capture the essential learning patterns without requiring the full training data to be stored or processed.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3970076B1Method for deep neural network compression
Publication Date: 2026.05.06 VOLKSWAGEN AG
  • EP3970076B1 patent drawingFigure 1
  • EP3970076B1 patent drawingFigure 2
  • EP3970076B1 patent drawingFigure 3

AI summary

A system and a method are provided for compressing a deep neural network ("DNN"). In some examples, the DNN is trained, where the DNN has at least one layer having multiple filters. Clustering of the filters of at least one layer is performed. Dimension reduction can be applied as well to the filters to reduce the channel dimensionality of the at least one layer. The dimensionally reduced DNN can then be retrained. Once retrained, the compressed DNN can be stored in a storage device.