Variable Bit-Width Quantization for Neural Network Image Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing techniques using artificial neural networks require large memory and complex circuits due to floating-point operations, and they do not effectively emulate human ocular focus perception, which prioritizes central vision over peripheral vision, leading to inefficient energy consumption and reduced accuracy in less important image regions.
Innovation Solution
Implementing quantization at different levels for image and weight data based on the perception characteristics of human ocular focus, where higher accuracy is maintained for central image regions and lower accuracy for peripheral regions, reducing memory requirements and energy consumption by using integer numbers with varying bit widths.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point numbers are used for weights and inputs in artificial neural network computations, then computation accuracy is maintained, but memory size requirements increase and circuit complexity increases
Solution Approach 1:
The patent applies parameter changes by transitioning from floating-point representation to fixed-point integer representation for weights and inputs. This changes the numerical parameter format, reducing the complexity of memory storage and circuit operations while maintaining sufficient computation accuracy for neural network inferences.
Solution Approach 2:
The patent implements different quantization levels for different regions of the image data. Central regions (fovea) use higher precision quantization to maintain accuracy where human vision is most sensitive, while peripheral regions use lower precision quantization, reducing overall computational complexity and memory requirements.
2Measurement precision
If floating-point numbers are used for weights and inputs in artificial neural network computations, then computation accuracy is maintained, but memory size requirements increase
Solution Approach 1:
The patent changes the parameter representation from floating-point to fixed-point integers with varying bit widths. This parameter transformation significantly reduces the memory size required to store weight and input data while preserving the essential computational accuracy needed for neural network operations.
Solution Approach 2:
Different quantization bit widths are applied to different regions: central image regions use higher bit widths (e.g., 8-bit) to maintain accuracy, while peripheral regions use lower bit widths (e.g., 4-bit or 2-bit), optimizing the balance between memory efficiency and computation accuracy.
3Measurement precision
If uniform high accuracy is applied to all image regions in artificial neural network processing, then overall image processing accuracy is maintained, but energy consumption increases
Solution Approach 1:
The patent applies local quality by differentiating processing accuracy based on spatial location. Central regions (corresponding to foveal vision) receive high-precision processing with more computational resources, while peripheral regions receive lower-precision processing, mimicking human visual attention and reducing overall energy consumption.
Solution Approach 2:
The image is segmented into different regions (central/foveal and peripheral) with different quantization requirements. This segmentation allows the system to allocate computational energy selectively, processing only critical regions at high accuracy while using reduced accuracy for less important regions.
4Device complexity
If quantization is applied to reduce memory requirements and simplify circuits, then device complexity is reduced, but computation accuracy deteriorates
Solution Approach 1:
The patent resolves this contradiction by applying local quality through region-specific quantization. Critical central regions maintain high computation accuracy with sufficient precision, while non-critical peripheral regions use lower precision, achieving overall system simplification without significant loss of functional accuracy.
Solution Approach 2:
The patent changes the numerical parameters from uniform floating-point to variable fixed-point representations, optimizing the precision level for each region's computational needs. This parameter optimization reduces circuit complexity while maintaining accuracy where it matters most.
Data Source
AI summary
A pair of smart glasses having: a digital camera configured to capture an image of a field of view; and a processing device configured to perform an analysis of the image using an artificial neural network having weight data. The processing device can apply different quantization levels to data from different regions of the image, and apply the different quantization levels to the weight data in weighing on the data from the different regions respectively. For example, weighing image data from a peripheral region of the image with the weight data can be performed with a lower level of accuracy than weighing image data from a center region of the image with the weight data to reduce energy consumption. Based on an output of the artificial neural network responsive to the image, the glasses can present virtual content superimposed on a view of reality seen through the glasses.