Demasking Engine for Pruned Neural Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional neural networks with large tensors are ill-suited for real-time applications due to slow processing speeds, and pruning them does not significantly accelerate inference time despite reducing network size.

Innovation Solution

A technique is introduced to remove masks from pruned neural networks by replacing masked tensors with denser versions, allowing for faster operations and propagating scatter operations to maintain output dimensionality, thereby optimizing neural network performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If pruning is applied to reduce network size, then memory footprint is reduced, but inference speed is not significantly accelerated

Engineering Contradiction:
Improvenetwork sizeVSAvoidinference speed
Core Design Contradiction:
Quantity of substanceVSProductivity

Solution Approach 1:

The patent extracts and removes the mask tensors from the pruned neural network, keeping only the active (non-zero) weights. This eliminates the computational overhead of processing zeroed-out elements while maintaining the reduced network size, thereby achieving both memory efficiency and speed improvement.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

Instead of keeping the full network and zeroing out elements via masks, the patent inverts the approach by directly representing only the active elements. This inversion transforms the data structure from a sparse masked tensor to a dense tensor containing only non-zero values, enabling faster computation without requiring mask application during inference.

Inventive Principle:
Principle #13The other way round (Inversion)

2Productivity

If masks are applied to zero out elements, then tensor operations become less meaningful, but removing masks restores original tensor dimensions

Engineering Contradiction:
Improveoperation efficiencyVSAvoidtensor dimensionality
Core Design Contradiction:
ProductivityVSLoss of information

Solution Approach 1:

The patent resolves the dimensionality issue by changing how the network parameters are organized. Instead of maintaining original tensor dimensions with zeroed elements, it reorganizes parameters into a compressed format that only includes active elements. The scatter operation then restores the expected output dimensions by placing results in appropriate positions, effectively managing dimensionality without sacrificing operational efficiency.

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

3Productivity

If dense tensors are used, then operations can be performed quickly, but memory consumption increases

Engineering Contradiction:
Improveprocessing speedVSAvoidmemory footprint
Core Design Contradiction:
ProductivityVSQuantity of substance

Solution Approach 1:

The patent applies local quality by treating different parts of the network differently - active weights are stored in a dense format for fast processing, while inactive weights are completely excluded from storage. This localized approach to density ensures that only the necessary portions of the network consume memory, achieving fast processing without proportional memory increase.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP3686816A1Techniques for removing masks from pruned neural networks
Publication Date: 2020.07.29 NVIDIA CORP
  • EP3686816A1 patent drawingFigure 1
  • EP3686816A1 patent drawingFigure 2
  • EP3686816A1 patent drawingFigure 3

AI summary

A demasking engine removes masks from a pruned neural network that is represented by a graph of nodes. The demasking engine analyzes a tensor and a mask associated with a given node in the graph of nodes to determine portions of the tensor that are zeroed by the mask. The demasking engine then removes these portions from the tensor to generate a densified tensor that has a smaller dimensionality than the original tensor. A function associated with the node can be evaluated more quickly based on the densified tensor than the original tensor. The demasking engine adds a scatter operation subsequent to the node in order to scale the dimensionality of the densified tensor to the dimensionality associated with the original tensor.