Deep Neural Network Compression With Skip Connections and Quantization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing deep neural networks require significant storage and computational resources, limiting their deployment in resource-constrained environments such as mobile and embedded systems, and existing compression methods often compromise network performance.

Innovation Solution

Introduce bridging connections within the neural network architecture, followed by parameter and activation quantization, to reduce resolution and storage requirements while maintaining performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If the resolution of parameters is reduced to decrease storage requirements, then storage space is saved, but network performance deteriorates

Engineering Contradiction:
Improvestorage spaceVSAvoidnetwork performance
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The neural network is segmented into multiple layers with different resolution requirements. The patent applies different quantization levels to different layers, allowing critical layers to maintain higher precision while non-critical layers use lower precision, thus preserving overall performance while reducing total storage requirements.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different parts of the neural network are assigned different quality levels (resolution). The patent identifies which layers benefit most from high precision and which can tolerate lower precision, applying local quality optimization to maintain performance where needed while reducing storage where possible.

Inventive Principle:
Principle #3Local quality

2Adaptability or versatility

If the resolution of parameters is reduced to enable deployment in resource-constrained environments, then adaptability improves, but computational precision deteriorates

Engineering Contradiction:
Improvedeployment capabilityVSAvoidcomputational precision
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent implements dynamic precision adjustment where the resolution of parameters can be adapted based on the specific deployment environment and task requirements. This allows the system to optimize between precision and resource usage dynamically rather than using a fixed precision level throughout.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the precision parameter of neural network parameters based on deployment constraints. By systematically varying the precision level across different layers and operations, the system adapts to resource-constrained environments while maintaining sufficient computational precision for accurate results.

Inventive Principle:
Principle #35Parameter changes

3Quantity of substance

If quantization is applied to reduce storage requirements, then storage efficiency improves, but information loss increases

Engineering Contradiction:
Improvestorage efficiencyVSAvoidparameter information
Core Design Contradiction:
Quantity of substanceVSLoss of information

Solution Approach 1:

The patent applies preliminary quantization analysis and layer importance assessment before final quantization. By pre-identifying which layers are most sensitive to quantization errors, the system can preserve more information in critical layers while applying more aggressive compression to non-critical layers, minimizing overall information loss.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent employs feedback mechanisms during the quantization process, using performance metrics to adjust quantization strategies. By monitoring the impact of quantization on network performance and adjusting precision levels accordingly, the system minimizes information loss while achieving storage efficiency goals.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3786853B1Compression of a deep neural network
Publication Date: 2025.09.10 ROBERT BOSCH GMBH
  • EP3786853B1 patent drawingFigure 0
  • EP3786853B1 patent drawingFigure 1
  • EP3786853B1 patent drawingFigure 2

AI summary

The invention relates to a method for compressing a deep neural network (12). The deep neural network comprises a plurality of layers, each of which is connected at its input to the immediately preceding layer in the sequence according to a predefinable order. The method comprises the following steps: adding a skip connection (24). This is followed by reducing the resolution of the layer parameters. The invention further relates to a computer program and a device for executing the method.