Neural Network Mask Propagation for Efficient Module Modification

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Modifying portions of a neural network requires significant memory, time, and computing resources due to the co-dependent nature of its components, making it challenging to determine which portions to modify and how to do so efficiently.

Innovation Solution

A system generates masks to modify neural network portions based on dependencies between network modules, using forward propagation and mask augmentation to identify and update neuron structures without backward propagation or gradient calculation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If traditional methods are used to modify neural network portions, then modification can be performed, but significant memory, time, and computing resources are required

Engineering Contradiction:

Improvemodification efficiencyVSAvoidcomputing resources

Core Design Contradiction:

ProductivityVSUse of energy by moving object

Solution Approach 1:

The neural network is divided into independent portions or modules, each with its own mask. When modifying one portion, only that specific segment needs to be updated rather than the entire network, significantly reducing computational resources and time required for modification.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Masks are pre-generated and stored for different portions of the neural network before modification is needed. These pre-computed masks enable rapid application to specific portions without requiring full network reprocessing, thus reducing time and computing resource consumption during actual modification operations.

Inventive Principle:

Principle #10Preliminary action

2Reliability

If all co-dependent portions are modified when one portion is changed, then network consistency is maintained, but modification time and resource usage increase

Engineering Contradiction:

Improvenetwork consistencyVSAvoidmodification time

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

Each portion of the neural network has its own mask that independently tracks modification requirements. This allows localized modification of only the affected portion without unnecessarily processing other independent portions, maintaining network consistency while reducing modification time.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

Masks act as intermediaries between modification operations and network portions. The masks capture dependency information and enable selective application of changes only to portions that actually depend on the modified section, avoiding redundant processing while ensuring consistency.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Manufacturing precision

If masks are generated for all portions of the neural network, then modification precision is improved, but memory requirements increase

Engineering Contradiction:

Improvemodification precisionVSAvoidmemory usage

Core Design Contradiction:

Manufacturing precisionVSQuantity of substance

Solution Approach 1:

Instead of generating masks for all possible portions, the system generates masks only for portions that are actually modified or needed. This partial action approach maintains modification precision for affected areas while significantly reducing overall memory consumption compared to pre-generating masks for the entire network.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS20250225394A1Modifying neural networks

Publication Date: 2025.07.10 NVIDIA CORP

AI summary

Apparatuses, systems, and techniques to generate one or more second masks to modify a second portion of a neural network based, at least in part, on one or more first masks of a first portion of the neural network from which the second portion of the neural network depends. In at least one embodiment, modifications one portion of a neural network are propagated to other portions of the neural network based on dependencies between these portions and one or more tensor masks.