Machine Learning Model Quantization Error Distribution

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Machine learning models are large and require significant storage and bandwidth, making them less appealing for use in bandwidth-limited networks and resource-constrained devices, such as mobile applications, where efficient distribution and storage are crucial.

Innovation Solution

The implementation of compression techniques, specifically model quantization with loss management, reduces the size of machine learning models by distributing quantization errors among associated weights, allowing for efficient distribution and storage while minimizing fitness degradation, along with strategies like patching, distributed learning, and transfer learning for improved model updates and usage.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If machine learning models are distributed in their original form, then model accuracy and performance are maintained, but storage capacity and transfer bandwidth requirements increase significantly

Engineering Contradiction:
Improvemodel accuracyVSAvoidstorage capacity
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent applies quantization by changing the precision parameter of model weights from high-precision floating-point representation to low-precision integer representation. This parameter change reduces the storage size of each weight while managing the introduction of quantization errors through compensation mechanisms that distribute errors across associated weights, thereby maintaining model accuracy despite reduced storage requirements

Inventive Principle:
Principle #35Parameter changes

2Reliability

If machine learning models are distributed in their original form, then model accuracy and performance are maintained, but transfer bandwidth requirements increase significantly

Engineering Contradiction:
Improvemodel accuracyVSAvoidtransfer bandwidth
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent reduces transfer bandwidth requirements by changing the data representation parameter from high-precision to low-precision format. The quantization process converts floating-point weights to integer weights, significantly reducing the number of bits that need to be transmitted over the network while implementing error compensation to preserve model performance

Inventive Principle:
Principle #35Parameter changes

3Reliability

If quantization error compensation is implemented by distributing errors among associated weights, then model fitness degradation is minimized, but computational complexity increases

Engineering Contradiction:
Improvemodel fitnessVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by pre-computing and storing the error compensation values during the quantization process. Instead of performing complex error distribution calculations during model inference, the compensation values are calculated in advance and embedded in the quantized model, reducing the computational burden during actual usage while maintaining model fitness

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3639206B1Systems and methods for compression and distribution of machine learning models
Publication Date: 2024.09.11 GOOGLE LLC
  • EP3639206B1 patent drawingFigure 1
  • EP3639206B1 patent drawingFigure 2
  • EP3639206B1 patent drawingFigure 3

AI summary

The present disclosure provides systems and methods for compressing and/or distributing machine learning models. In one example, a computer-implemented method is provided to compress machine-learned models, which includes obtaining, by one or more computing devices, a machine-learned model. The method includes selecting, by the one or more computing devices, a weight to be quantized and quantizing, by the one or more computing devices, the weight. The method includes propagating, by the one or more computing devices, at least a part of a quantization error to one or more non-quantized weights and quantizing, by the one or more computing devices, one or more of the non-quantized weights. The method includes providing, by the one or more computing devices, a quantized machine-learned model.