Depth-Wise Convolution Initialization for Stable Neural Network Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The initialization of neural network parameters significantly impacts training efficiency, with unsuitable initializations leading to slow convergence or failure in training, particularly in depth-wise convolutional layers where the number of inputs and outputs differ substantially.

Innovation Solution

Initializing the parameters of depth-wise convolutional layers using a predefined probability distribution with variance characterized by the reciprocal of the square root of the product of the number of inputs and outputs, ensuring gradients are neither too small nor too large, thereby facilitating faster convergence.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If parameters are initialized with standard random distribution, then training can proceed, but convergence is slow or fails when number of inputs and outputs differ substantially

Engineering Contradiction:
Improvetraining convergenceVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent modifies the initialization parameters by setting the variance of the random distribution to the reciprocal of the square root of the product of the number of inputs and outputs. This parameter change ensures that gradients remain within an appropriate range during backpropagation, preventing both vanishing and exploding gradient problems that occur with standard initialization methods when input and output dimensions differ substantially.

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If number of inputs and outputs in depth-wise convolutional layer differ substantially, then layer can handle complex transformations, but standard initialization causes gradients to be too small or too large

Engineering Contradiction:
Improvelayer transformation capabilityVSAvoidgradient stability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent changes the initialization parameter (variance) to be inversely proportional to the square root of the product of input and output counts. This adjustment compensates for the dimensional mismatch in depth-wise convolutional layers, ensuring that the scale of gradients remains stable regardless of how differently the number of inputs and outputs are scaled.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4258177A1Method for initializing a neural network
Publication Date: 2023.10.11 ROBERT BOSCH GMBH
  • EP4258177A1 patent drawingFigure 1
  • EP4258177A1 patent drawingFigure 2
  • EP4258177A1 patent drawingFigure 3

AI summary

Computer-implemented method (100) for training a neural network (60) wherein the neural network (60) is configured to determine an output signal (y) based on an input signal (x) and wherein training comprises training parameters of a depth-wise convolutional layer of the neural network (60), wherein the depth-wise convolutional layer is initialized based on values drawn from a predefined probability distribution, wherein a variance of the probability distribution is characterized by a reciprocal of a square root of a number of filters applied at each depth of an input of the depth-wise convolutional layer.