Neural Network Pruning Mask Optimization for Memory Reduction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Artificial neural networks face challenges in maintaining performance while reducing complexity to avoid overfitting and excessive memory allocation, which hinders miniaturization and commercialization.

Innovation Solution

A method and apparatus that determine a pruning mask to maximize weight importance in a trained neural network by using a processor to express weight importance as a pruning binary vector and apply constraints, optimizing weight pruning based on binary vector optimization algorithms to reduce complexity.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If the complexity of a neural network is increased to improve learning capacity and accuracy, then the accuracy of learned data increases, but confidence in estimated value of newly learned data decreases and overfitting occurs

Engineering Contradiction:
Improveaccuracy of learned dataVSAvoidconfidence in estimated value
Core Design Contradiction:
Manufacturing precisionVSReliability

Solution Approach 1:

The patent extracts and removes redundant or less important weights from the neural network through pruning operations. By identifying and eliminating unnecessary connections between neurons, the network complexity is reduced while maintaining the essential learning capacity, thus preventing overfitting and improving generalization to new data.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the parameters of the neural network by adjusting the pruning threshold and pruning ratio. These parameter changes allow the network to transition between different complexity levels, optimizing the balance between accuracy and reliability for different application scenarios.

Inventive Principle:
Principle #35Parameter changes

2Manufacturing precision

If the complexity of a neural network is increased to improve learning capacity, then the accuracy increases, but memory allocation excessively increases

Engineering Contradiction:
ImproveaccuracyVSAvoidmemory allocation
Core Design Contradiction:
Manufacturing precisionVSQuantity of substance

Solution Approach 1:

The patent removes unnecessary weights and connections from the neural network through systematic pruning. This extraction process reduces the quantity of stored parameters while preserving the network's accuracy on training data, directly addressing the memory allocation issue.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies partial pruning by removing only a portion of the weights rather than all connections. By carefully selecting which weights to prune based on importance metrics, the network maintains sufficient complexity for accurate learning while reducing memory requirements to practical levels.

Inventive Principle:
Principle #16Partial or excessive action

3Manufacturing precision

If the complexity of a neural network is increased to improve learning capacity, then the accuracy increases, but miniaturization and commercialization are hindered

Engineering Contradiction:
ImproveaccuracyVSAvoidminiaturization
Core Design Contradiction:
Manufacturing precisionVSEase of manufacture

Solution Approach 1:

The patent extracts redundant computational elements from the neural network, reducing the overall size and complexity. This enables the network to be implemented in smaller, more cost-effective hardware configurations suitable for miniaturization and commercial deployment.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent modifies the network architecture parameters through pruning to create a compact version that maintains accuracy while being suitable for smaller hardware platforms. This facilitates ease of manufacture and commercialization by adapting the network to practical deployment constraints.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4231203A1Method and apparatus with pruning background
Publication Date: 2023.08.23 SAMSUNG ELECTRONICS CO LTD

AI summary

A method and apparatus with pruning are disclosed. A method is performed by an apparatus including a processor, the method includes determining weight importance of a trained neural network, receiving a constraint condition related to an operation resource, and determining, in accordance with the constraint condition, a pruning mask for maximizing the weight importance of the trained neural network.