MoE Layer Placement in Neural Networks Using Surrogate Layers

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Constructing a single monolithic neural network architecture for diverse sensor data applications is computationally expensive and inefficient, as it requires extensive computational resources and training complexity, especially when integrating Mixture of Experts (MoE) functionality.

Innovation Solution

A method to determine optimal layers for integrating MoE functionality by training surrogate layers, which are computationally cheaper and provide performance upper bounds, allowing for efficient resource allocation and reduced training complexity.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a monolithic neural network architecture is used to handle diverse sensor data, then the network can process all situations, but the computational cost and training complexity increase significantly

Engineering Contradiction:
Improveability to handle diverse sensor data situationsVSAvoidcomputational cost
Core Design Contradiction:
Adaptability or versatilityVSPower

Solution Approach 1:

The patent divides the monolithic neural network into multiple specialized processing blocks (experts), each handling specific types of sensor data or situations. A router block segments incoming data and directs it to appropriate experts, reducing computational load while maintaining versatility across diverse applications like image classification, object detection, and sensor fusion.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The Mixture of Experts architecture creates a universal processing framework where multiple specialized blocks work together to handle diverse sensor data types. The router block provides multi-functionality by dynamically selecting appropriate experts based on input characteristics, enabling a single architecture to serve multiple applications including autonomous vehicles, medical imaging, and industrial inspection.

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

2Productivity

If Mixture of Experts functionality is integrated into all layers of the neural network, then processing efficiency improves, but device complexity and training difficulty increase

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidnetwork architecture complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent applies MoE functionality selectively to specific layers or blocks where it provides the most benefit, rather than uniformly across all layers. This local application optimizes processing efficiency for critical operations while avoiding unnecessary complexity in layers where standard processing suffices, balancing productivity gains with architectural manageability.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The implementation uses partial MoE action by activating only a subset of expert blocks for each input based on router decisions. This partial activation provides efficiency improvements without requiring all expert blocks to be fully trained and deployed, reducing the effective complexity while maintaining productivity benefits for relevant data types.

Inventive Principle:
Principle #16Partial or excessive action

3Measurement precision

If the number of parameters in deep learning models is increased to improve performance, then model accuracy improves, but inference time remains compromised

Engineering Contradiction:
Improvemodel accuracyVSAvoidinference time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent segments the large parameter set across multiple expert blocks, each with specialized parameters for specific data types. The router blocks segment the inference process by directing inputs to only the relevant experts, allowing high accuracy through specialized parameters while maintaining fast inference by avoiding computation through inactive blocks.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system uses partial action by activating only the necessary subset of expert blocks for each inference based on input characteristics. This allows the model to leverage large parameter counts for accuracy when needed, while maintaining fast inference times by skipping computation through blocks whose parameters are not required for the current input type.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4687070A1Optimizing mixture of experts, moe, integration into neural network architectures
Publication Date: 2026.02.04 ROBERT BOSCH GMBH
  • EP4687070A1 patent drawingFigure 1
  • EP4687070A1 patent drawingFigure 2
  • EP4687070A1 patent drawingFigure 3

AI summary

A method (100) for determining, in a neural network (1) that comprises N layers (1a-1d) and is configured for classification and/or regression of sensor data (2), one or more optimal layers (1*) for the integration of Mixture of Experts, MoE, functionality, said MoE functionality comprising a plurality of M distinct processing blocks (4a-4d) and a router block (3) that routes each input (I) to one or more processing blocks (4a-4d) for processing, the method (100) comprising the steps of: • constructing (110) candidate versions (1A, 1B) of the neural network (1), wherein, in each candidate version (1A, 1B), one or more layers (1a-1d) are replaced with surrogate layers (1a'-1d'), wherein: ∘ each surrogate layer (1a'-1d') is computationally cheaper to train than a respective MoE layer with full MoE functionality, while ∘ the performance of the candidate neural network (1A, 1B) with the surrogate layer (1a'-1d') is commensurate with the performance that the neural network (1) would have with the MoE layer in the place of the surrogate layer (1a'-1d'); • training (120), using training examples (2a) of sensor data, each candidate version (1A, 1B) of the neural network (1); • determining (130), using test and/or validation samples (2b) of sensor data for which respective ground truth outputs (5b) of the neural network (1) are known, the accuracy (6) with which the trained candidate version (1A*, 1 B*) of the neural network (1) reproduces the ground truth outputs (5b); and • determining (140) the layers (1a-1d) that are replaced with surrogate layers in a candidate version (1A, 1B) of the neural network (1) with a best accuracy (6) as optimal layers (1*) for the integration of MoE functionality.