Neural Network Weight Quantization for Analog MAC Accuracy

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing quantization methods for analog computing in neural networks are inefficient and lack accuracy, particularly in converting floating-point weights to fixed-point or integer types for non-volatile memory arrays, which affects energy consumption and computational efficiency.

Innovation Solution

A method for quantizing weights in analog computing that involves dividing weights into multiple quantization unit groups within each layer, performing MAC operations, and dequantizing the results, using non-volatile memory elements and converters to improve accuracy and reduce energy consumption.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If weights are quantized from floating-point to fixed-point or integer types for analog computing, then energy consumption is reduced and computational efficiency is improved, but quantization errors increase and accuracy decreases

Engineering Contradiction:

Improveenergy consumptionVSAvoidquantization accuracy

Core Design Contradiction:

Use of energy by moving objectVSMeasurement precision

Solution Approach 1:

The patent divides weights into multiple quantization unit groups, where each group is quantized separately with its own quantization parameters. This segmentation allows different precision levels for different weight groups, reducing overall quantization error while maintaining energy efficiency in analog computing

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent applies different quantization strategies to different weight groups based on their specific characteristics. Each quantization unit group can have customized quantization parameters, enabling local optimization of precision where needed while accepting lower precision where tolerable, thus balancing accuracy and energy consumption

Inventive Principle:

Principle #3Local quality

2Ease of manufacture

If a single quantization method is applied to all weights in a layer, then device complexity is reduced and ease of operation is improved, but manufacturing precision and accuracy of analog computing deteriorate

Engineering Contradiction:

Improvequantization implementation simplicityVSAvoidweight quantization precision

Core Design Contradiction:

Ease of manufactureVSManufacturing precision

Solution Approach 1:

The patent segments weights into multiple quantization unit groups that can be processed independently. This segmentation enables sophisticated quantization strategies without requiring complex global optimization, as each group can be quantized separately with simpler local methods

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary classification of weights into quantization unit groups based on their characteristics before quantization. This preliminary action allows the system to prepare appropriate quantization parameters in advance, simplifying the actual quantization process while ensuring high precision

Inventive Principle:

Principle #10Preliminary action

3Measurement precision

If more quantization unit groups are used within each layer, then quantization accuracy is improved and energy consumption is reduced, but device complexity and computational overhead increase

Engineering Contradiction:

Improvequantization accuracyVSAvoidquantization unit group management complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent organizes weights into multiple quantization unit groups that map to specific memory regions and processing units. This segmentation enables parallel processing of different groups, improving accuracy through finer-grained quantization while distributing complexity across multiple independent processing paths

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent designs a universal quantization framework where the same basic quantization mechanism can be applied to multiple different quantization unit groups. This multi-functionality allows the system to handle complex multi-group quantization scenarios using a standardized approach, reducing overall system complexity

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS20250348716A1Weight quantization method for analog computing of neural network model and device for performing the same

Publication Date: 2025.11.13 IHW INC

AI summary

An analog computing method for storing weights in non-volatile memory elements arranged in a memory array and performing a multiply-accumulate calculation (MAC) operation includes a quantization step for converting the weights, which are included in each of a plurality of layers for operations in a neural network model including the layers, from first weights represented in floating-point numbers to second weights by quantizing the first weights to fixed-point numbers or integers, a weight storage step in which the second weights are stored in the non-volatile memory elements arranged in the memory array, a MAC operation step in which an input signal is applied to the memory array to perform the MAC operation to output a MAC operation result, a digital conversion step in which the output MAC operation result is converted into a digital MAC operation result that is a digital signal.