Conditional Convolutional Layers With Soft Kernel Selection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional convolutional neural networks face challenges in increasing model capacity without significant computational overhead, and existing conditional computation models are difficult to train due to discrete routing requirements.

Innovation Solution

Implementing conditional convolutional layers with input-dependent kernel weights and soft kernel selection, which allows for increased computational capacity with minimal additional processing power, avoiding discrete routing decisions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If conventional convolutional layers are used, then computational overhead is low, but model capacity and performance are limited

Engineering Contradiction:
Improvemodel capacityVSAvoidcomputational overhead
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements dynamic kernel selection where the network adaptively chooses which kernels to apply based on the input features. This is achieved through learnable routing weights that dynamically weight different kernel sets, allowing the model capacity to adapt to different inputs while keeping the average computational overhead controlled through sparse activation.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the parameters of the convolutional layer by introducing input-dependent kernel selection. Instead of using fixed kernels, the system learns to select and combine different kernel sets based on the input, effectively changing the operational parameters dynamically. This is implemented through learnable routing parameters that determine kernel activation.

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If conditional computation models are used to increase model capacity, then performance improves, but training becomes difficult due to discrete routing decisions

Engineering Contradiction:
Improvemodel capacityVSAvoidtraining difficulty
Core Design Contradiction:
Adaptability or versatilityVSEase of manufacture

Solution Approach 1:

The patent replaces the discrete routing mechanism (which is difficult to optimize) with a continuous, differentiable routing mechanism. Instead of making discrete decisions about which kernels to activate, the system uses learnable continuous weights to smoothly combine multiple kernel sets, enabling gradient-based optimization while maintaining conditional computation benefits.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent implements dynamic kernel selection where the network adaptively chooses which kernels to apply based on the input features. This is achieved through learnable routing weights that dynamically weight different kernel sets, allowing the model capacity to adapt to different inputs while keeping the average computational overhead controlled through sparse activation.

Inventive Principle:
Principle #15Dynamics

3Measurement precision

If the neural network size is increased to improve performance, then accuracy improves, but resource consumption increases significantly

Engineering Contradiction:
ImproveaccuracyVSAvoidresource consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies local quality by having different regions of the kernel space specialized for different features, and by selectively activating only the relevant kernels for each input. This allows the network to have large model capacity locally (through multiple kernel sets) while consuming fewer resources globally (through sparse activation), improving accuracy without proportional resource consumption.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP3899806B1Convolutional neural networks with soft kernel selection
Publication Date: 2025.12.10 GOOGLE LLC
  • EP3899806B1 patent drawingFigure 1
  • EP3899806B1 patent drawingFigure 2~3
  • EP3899806B1 patent drawingFigure 4

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing inputs using neural networks that include one or more conditional convolutional layers. A conditional convolutional layer has a plurality of kernels and determines a respective input-dependent weight for each of the plurality of kernels and generates an input-dependent kernel by computing a weighted sum of the plurality of kernels in accordance with the respective input-dependent weights.