Neural Network Training via Second Network Gradient Approximation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The existing neural network training methods require complex calculations for second-order differentials to adjust gradients, which can be computationally intensive due to the need for Hessian matrix calculations.

Innovation Solution

A method involving a first neural network to estimate output data and a second neural network to generate output differential values with respect to input data, where the second neural network differentiates the output of each layer of the first neural network using shared parameters, allowing for training without direct calculation of second-order differentials.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If second-order differential calculations are performed to calculate gradients for neural network training, then training accuracy can be improved, but computational complexity and calculation time increase significantly

Engineering Contradiction:
Improvetraining accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent introduces a second neural network as an intermediary component that approximates the second-order differential calculations. Instead of directly computing complex second-order derivatives, the system uses the second neural network to generate differential values based on input data and first-order differential data from the first neural network, significantly reducing computational complexity while maintaining training accuracy

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent creates a copy of the neural network architecture (the second neural network) that replicates the computational function of calculating second-order differentials. This copied network takes first-order differential data as input and produces the required differential values, avoiding the need for direct second-order differentiation while preserving the necessary computational functionality

Inventive Principle:
Principle #26Copying

2Measurement precision

If Hessian matrix calculations are performed to obtain second-order differentials, then gradient accuracy improves, but processing time and energy consumption increase

Engineering Contradiction:
Improvegradient accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent replaces the mechanical computation of Hessian matrices with a neural network-based system. The second neural network substitutes for the complex mathematical operations of second-order differentiation, enabling gradient accuracy to be maintained while processing time is significantly reduced through learned patterns rather than brute-force calculation

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

3Reliability

If second-order differential calculations are performed for all input data, then training completeness is ensured, but computational efficiency decreases

Engineering Contradiction:
Improvetraining completenessVSAvoidcomputational efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent applies partial action by determining whether second-order differential calculations are necessary for each input data sample. The system uses the second neural network to generate differential values selectively, avoiding unnecessary computations for inputs where first-order differentials suffice, thus maintaining training completeness while improving computational efficiency

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4401009A1Method and apparatus with neural network training
Publication Date: 2024.07.17 SAMSUNG ELECTRONICS CO LTD
  • EP4401009A1 patent drawingFigure 1
  • EP4401009A1 patent drawingFigure 2
  • EP4401009A1 patent drawingFigure 3

AI summary

A processor-implemented method may include generating respective first neural network differential data by differentiating a respective output of each layer of a first neural network with respect to input data provided to the first neural network that estimates output data from the input data, by a forward propagation process of the first neural network, generating, using a second neural network, an output differential value of the output data with respect to the input data using the respective first neural network differential data, and training the first neural network and the second neural network based on ground truth data of the output data and ground truth data of the output differential value.