Factorized Pointwise Convolution for Faster CNN Channel Fusion

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional pointwise convolutions in machine learning models, particularly in deep neural networks, suffer from high computational complexity and resource requirements, especially in channel fusion operations, which can be a bottleneck for efficient inference processing.

Innovation Solution

Implement fast pointwise convolution methods that utilize factorization and reduced weight density to reduce the number of multiplications and trainable weights, achieving a complexity of N * log N + N, thereby improving computational efficiency and reducing memory usage.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If standard convolution operations are used in neural network layers, then comprehensive feature extraction is achieved, but computational complexity and processing time increase significantly

Engineering Contradiction:
Improvefeature extraction accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent segments the convolution operation into two distinct parts: a shared weight matrix computation that can be pre-calculated and stored, and a pointwise multiplication operation that is computationally efficient. This segmentation allows the complex convolution to be broken down into manageable components that can be optimized independently, reducing overall processing time while maintaining feature extraction accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary computation of the weight matrix by pre-calculating the outer product of input features and filter weights during the training phase or before inference. This pre-computed weight matrix is then stored and reused during inference, eliminating the need to perform complex multiplications during real-time processing, thus significantly reducing inference time.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If multiple convolution filters are applied to extract diverse features, then feature representation quality improves, but computational resources and processing overhead increase

Engineering Contradiction:
Improvefeature representation qualityVSAvoidcomputational resources
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges multiple convolution operations into a single efficient computation by combining multiple filters into one composite filter. Instead of applying each filter separately and then combining results, the method computes a single weight matrix that encapsulates the combined effect of multiple filters, reducing computational overhead while preserving the ability to extract diverse features.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The shared weight matrix serves multiple purposes: it can be used for different input sequences, different positions in the sequence, and can represent multiple filter operations simultaneously. This universal computation structure eliminates the need for separate computational paths for each filter or position, reducing device complexity while maintaining comprehensive feature extraction capabilities.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Reliability

If sequential processing of input sequences is used, then temporal dependencies are captured, but processing speed decreases

Engineering Contradiction:
Improvetemporal dependency captureVSAvoidprocessing speed
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent pre-computes the weight matrix that captures temporal relationships between sequence elements before actual inference. By performing this computation in advance, the system can quickly apply the pre-computed weights to new input sequences without performing expensive sequential computations during inference, thus maintaining temporal dependency capture while improving processing speed.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses the pre-computed weight matrix as a template that can be copied and applied to multiple different input sequences. Instead of recomputing temporal relationships for each sequence, the system copies the pre-established weight matrix and applies it to new data, dramatically improving processing speed while preserving the captured temporal dependencies.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4032028B1Efficient inferencing with fast pointwise convolution
Publication Date: 2026.05.20 QUALCOMM INC
  • EP4032028B1 patent drawingFigure 1A~1C
  • EP4032028B1 patent drawingFigure 1D
  • EP4032028B1 patent drawingFigure 2

AI summary

Embodiments described herein relate to a method, comprising: receiving input data at a convolutional neural network (CNN) model; generating a factorized computation network comprising a plurality of connections between a first layer of the CNN model and a second layer of the CNN model, wherein: the factorized computation network comprises N inputs, the factorized computation network comprises M outputs, and the factorized computation network comprises at least one path from every input of the N inputs to every output of the M outputs; setting a connection weight for a plurality of connections in the factorized computation network to 1 so that a weight density for the factorized computation network is < 100%; performing fast pointwise convolution using the factorized computation network to generate fast pointwise convolution output; and providing the fast pointwise convolution output to the second layer of the CNN model.