Compound Coefficient Scaling for Resource-Aware Neural Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional methods for scaling neural networks are inefficient and computationally expensive due to the need for extensive hyper-parameter tuning, often resulting in low accuracy and inefficiency when increasing network width, depth, or resolution.

Innovation Solution

A system that determines a final neural network architecture by uniformly scaling network width, depth, and resolution using a compound coefficient, optimizing resource usage and maximizing model accuracy without extensive hyper-parameter tuning.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If network width, depth, or resolution is increased to improve model accuracy, then model accuracy is improved, but computational cost and hyper-parameter tuning complexity increase significantly

Engineering Contradiction:
Improvemodel accuracyVSAvoidhyper-parameter tuning complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies parameter changes by introducing a compound scaling coefficient that simultaneously adjusts multiple architectural parameters (width, depth, resolution) based on available computational resources. This transforms the complex multi-dimensional hyper-parameter tuning problem into a single-parameter optimization, where the compound coefficient encapsulates the relationships between width, depth, and resolution scaling factors.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The compound scaling coefficient serves multiple functions simultaneously: it scales network width, adjusts network depth, and modifies resolution based on resource constraints. This single parameter performs what would traditionally require separate tuning of multiple independent hyper-parameters, reducing complexity while maintaining the ability to optimize all critical architectural dimensions.

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

2Measurement precision

If network width, depth, or resolution is increased to improve model accuracy, then model accuracy is improved, but computational resources and training time increase

Engineering Contradiction:
Improvemodel accuracyVSAvoidcomputational resources
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent uses parameter changes to establish a resource-aware scaling mechanism where the compound coefficient is derived from computational resource constraints. This ensures that model accuracy improvements are achieved at controlled computational costs by mathematically linking architectural scaling to available FLOPs, memory, and other resource metrics.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The scaling approach is dynamic and adaptive to resource availability. Rather than using fixed scaling rules, the compound coefficient adjusts the scaling behavior based on the specific computational resources provided, allowing the model to automatically optimize the trade-off between accuracy and computational cost for different deployment scenarios.

Inventive Principle:
Principle #15Dynamics

3Quantity of substance

If conventional scaling methods are used to increase model size, then model capacity increases, but efficiency and accuracy deteriorate due to arbitrary scaling

Engineering Contradiction:
Improvemodel capacityVSAvoidtraining and inference efficiency
Core Design Contradiction:
Quantity of substanceVSProductivity

Solution Approach 1:

The patent transforms arbitrary scaling into principled scaling by changing the parameter relationships. Instead of independently adjusting width, depth, and resolution, the compound coefficient enforces coordinated changes that maintain optimal architectural proportions. This ensures model capacity increases while preserving efficiency by maintaining balanced growth across all architectural dimensions.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3912099B1Compound model scaling for neural networks
Publication Date: 2025.10.15 GOOGLE LLC
  • EP3912099B1 patent drawingFigure 1
  • EP3912099B1 patent drawingFigure 2
  • EP3912099B1 patent drawingFigure 3

AI summary

A method for determining a final architecture for a neural network to perform a particular machine learning task is described. The method includes receiving a baseline architecture for the neural network, wherein the baseline architecture has a network width dimension, a network depth dimension, and a resolution dimension; receiving data defining a compound coefficient that controls extra computational resources used for scaling the baseline architecture; performing a search to determine a baseline width, depth and resolution coefficient that specify how to assign the extra computational resources to the network width, depth and resolution dimensions of the baseline architecture, respectively; determining a width, depth and resolution coefficient based on the baseline width, depth, and resolution coefficient and the compound coefficient; and generating the final architecture that scales the network width, network depth, and resolution dimensions of the baseline architecture based on the corresponding width, depth, and resolution coefficients.