Neural Network Configuration Search Optimization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Finding suitable neural network configurations that are optimized for prediction error and additional optimization goals, such as hardware limitations, is time-consuming due to the immense number of possible configurations and the need for extensive training and evaluation.

Innovation Solution

A method to determine a Pareto set of optimized network configurations using a probability distribution for preselecting network configuration candidates, iteratively updating the set based on prediction errors and resource costs, and selecting variants that improve the Pareto front, allowing for a resource-saving approach to find suitable configurations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of time

If the number of candidate network configurations is reduced using expert knowledge, then the search time is reduced, but the quality and optimality of the selected configuration may deteriorate

Engineering Contradiction:
Improvesearch timeVSAvoidconfiguration quality
Core Design Contradiction:
Loss of timeVSMeasurement precision

Solution Approach 1:

The patent applies preliminary action by pre-selecting a subset of promising network configurations from the search space before full training and evaluation. This pre-selection is based on architectural heuristics and performance estimates, allowing the system to focus computational resources on the most promising candidates while still maintaining the possibility of finding high-quality configurations.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent changes parameters by using multiple performance metrics (including accuracy, training time, and model size) to evaluate and filter configurations. By adjusting the weighting and thresholds of these parameters, the system can balance between search efficiency and configuration quality, finding optimal configurations without exhaustively searching the entire space.

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If multiple optimization goals (prediction error and hardware limitations) are considered simultaneously, then the applicability to practical systems is improved, but the complexity of the search process increases

Engineering Contradiction:
Improveapplicability to practical systemsVSAvoidsearch process complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent segments the optimization process into distinct phases: initial configuration generation, pre-selection based on architectural criteria, training evaluation, and final filtering based on multiple optimization goals. This segmentation allows each phase to focus on specific criteria, managing the overall complexity while achieving multi-objective optimization.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent adds another dimension to the search space by incorporating hardware-specific parameters (model size, training time, computational requirements) alongside traditional accuracy metrics. This multi-dimensional evaluation framework enables simultaneous optimization for both prediction error and hardware limitations without exponentially increasing search complexity.

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

3Measurement precision

If each candidate network configuration is fully trained and evaluated, then the accuracy of performance assessment is improved, but the computational cost and search time increase significantly

Engineering Contradiction:
Improveperformance assessment accuracyVSAvoidcomputational cost
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies partial action by training and fully evaluating only a subset of the most promising network configurations rather than all possible configurations. This approach performs sufficient training on selected candidates to obtain accurate performance assessments while avoiding the excessive computational cost of evaluating the entire search space.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent uses preliminary filtering based on architectural heuristics and performance estimates to identify promising configurations before committing computational resources to full training. This preliminary action ensures that accurate performance assessment is applied only to configurations with high potential, reducing overall computational cost while maintaining assessment accuracy for the selected candidates.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3785177B1Method and device for determining a network-configuration of a neural network
Publication Date: 2023.07.05 ROBERT BOSCH GMBH
  • EP3785177B1 patent drawingFigure 1~2
  • EP3785177B1 patent drawingFigure 3
  • EP3785177B1 patent drawingFigure 4

AI summary

The invention relates to methods for determining a suitable network configuration for a neural network for a predefined application, which is determined in the form of training data, wherein the network configuration specifies the architecture of the neural network, having the following steps: a) providing (S1) a current network configuration set having network configurations, wherein the current network configuration set corresponds to a Pareto set with regard to a prediction error and at least one further optimisation target; b) providing (S2) a set of network configuration variants; c) selecting network configurations (S3) from the set of network configuration variants, depending on a probability distribution of the network configurations of the current network configuration set with respect to the at least one further optimisation target; d) training (S4) neural nets of each of the selected network configurations and determining a corresponding prediction error (error(N); e) updating (S5) the current network configuration set on the basis of the prediction errors and the at least one further optimisation target of the network configuration set and the selected network configurations; and f) selecting the suitable network configuration from the current network configuration set.