Adaptive Bit-Width Neural Network Quantization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep learning models, such as CNNs, are large and require significant computing resources, making them impractical for real-time applications on portable devices due to limited battery life, memory, and processing power, with existing solutions either compromising model accuracy through model reduction or being limited by hardware upgrade costs and energy concerns.

Innovation Solution

A method for generating high-accuracy, lightweight machine learning models with adaptive bit-widths for different layers, using a combination of network pruning and adaptive quantization techniques, allowing for reduced model size without substantial loss in accuracy, even without integer weight regularization and 8-bit forward quantization during training.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If deep learning models are deployed for real-time applications on portable devices, then model accuracy is improved, but device complexity and resource consumption increase

Engineering Contradiction:
Improvemodel accuracyVSAvoidcomputational footprint
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies parameter changes by systematically varying bit-width parameters across different network layers during training. Each layer is assigned a specific bit-width (e.g., 8-bit, 6-bit, 4-bit) to control the precision of weight representations. This parameter transformation enables the model to maintain accuracy while reducing computational complexity and memory requirements, directly resolving the contradiction between model accuracy and device complexity.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent implements local quality by applying different bit-width quantization levels to different layers of the neural network based on their specific requirements. Critical layers maintain higher precision (e.g., 8-bit) while less sensitive layers use lower precision (e.g., 4-bit). This localized approach ensures that accuracy is preserved where needed while reducing overall computational footprint, effectively balancing the contradiction between reliability and device complexity.

Inventive Principle:
Principle #3Local quality

2Device complexity

If model reduction techniques are applied to reduce computational footprint, then device complexity is reduced, but model accuracy deteriorates

Engineering Contradiction:
Improvecomputational footprintVSAvoidmodel accuracy
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The patent employs dynamics by making the quantization bit-width configurable and adaptable rather than fixed. The system can dynamically adjust the bit-width assignment for different layers based on performance requirements and resource constraints. During training, the model learns optimal representations at each bit-width level, enabling accurate inference even with reduced precision. This dynamic approach allows the model to maintain high accuracy while achieving significant reduction in computational footprint.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent applies preliminary action by pre-training the neural network with quantized weights at various bit-width levels before deployment. This preliminary quantization training ensures that the model adapts to lower precision representations in advance, so that when deployed with reduced computational footprint, it maintains accuracy comparable to full-precision models. This pre-adaptation prevents accuracy deterioration that would normally result from simple model reduction.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If hardware upgrades are implemented to support large models, then model accuracy is improved, but energy consumption and cost increase

Engineering Contradiction:
Improvemodel accuracyVSAvoidenergy consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent substitutes mechanical/hardware solutions with a software-based quantization approach. Instead of upgrading hardware to support large models, the system transforms the model representation to use lower bit-width weights that can be efficiently processed by existing hardware. This substitution of quantization techniques for hardware upgrades achieves high accuracy while maintaining low energy consumption, as the reduced precision enables faster computation and lower power usage on portable devices.

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

Data Source

PatentEP3619652B1Adaptive bit-width reduction for neural networks
Publication Date: 2021.11.24 MIDEA GROUP CO LTD
  • EP3619652B1 patent drawingFigure 1
  • EP3619652B1 patent drawingFigure 2
  • EP3619652B1 patent drawingFigure 3

AI summary

A method of providing an adaptive bit-width neural network model on a computing device, comprising: obtaining a first neural network model, wherein each layer of first neural network model has a respective set of parameters expressed with an original bit-width of the first neural network model; reducing a footprint of the first neural network model by using respective reduced bit-widths for storing the respective sets of parameters of different layers of the first neural network model, wherein: preferred values of the respective reduced bit-widths are determined through multiple iterations of forward propagation through the first neural network model using a validation data set while each of two or more layers of the first neural network model is expressed with different degrees of quantization until a predefined information loss threshold is met; and generating a reduced neural network model with quantized parameters expressed with the respective reduced bit-widths.