Real Signal Quantization Using Distance-Encoded Nonzero Bits
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing signal processing technologies face challenges in efficiently quantizing real signals for neural networks and other applications, requiring significant computational resources and memory, while maintaining precision and performance.
Innovation Solution
A method for coding real signals into quantized signals by converting them into binarized samples with selected significant bits, using binary or canonical signed binary representations, and encoding distances between bits to minimize resource usage, with optional decomposition of high-value distances for efficient multiplication and convolution operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If real signals are quantized using traditional methods, then computational resources and memory requirements are reduced, but precision and accuracy of signal processing are degraded
Solution Approach 1:
The patent segments the real signal into two distinct components: a sign bit and a magnitude represented by distances between non-zero bits. This segmentation allows the magnitude to be encoded with fewer bits while preserving the relative positional information, thus reducing memory requirements without proportionally degrading precision. The sign bit separately captures the polarity information, and the distance encoding captures the magnitude information in a compressed form.
Solution Approach 2:
The patent transforms the traditional magnitude representation into a different dimensional space by encoding distances between non-zero bits rather than directly encoding the magnitude value. This dimensional transformation allows the same information to be represented with fewer bits, as the distance between bits provides a logarithmic-like compression of the magnitude information while preserving the relative precision needed for signal processing operations.
2Measurement precision
If more bits are used for quantization, then signal precision is maintained, but computational complexity and resource requirements increase
Solution Approach 1:
The patent segments the quantization process into sign extraction and magnitude encoding, where the magnitude is represented by distances between non-zero bits. This segmentation reduces the total number of bits needed while maintaining precision, as the distance encoding efficiently captures magnitude information without requiring full precision representation of every possible value.
Solution Approach 2:
The patent changes the parameter representation from direct magnitude encoding to distance-based encoding. By representing the magnitude as the distance between non-zero bits rather than the absolute value, the system achieves the same precision with fewer bits, thereby reducing computational complexity and resource requirements while maintaining the necessary signal fidelity.
3Ease of manufacture
If traditional quantization methods are used, then implementation is straightforward, but multiplication and convolution operations remain complex
Solution Approach 1:
The patent segments numbers into sign bits and distance-encoded magnitudes, which enables simplified multiplication and convolution operations. When multiplying two numbers represented this way, the system can work with the distance values and sign bits separately, avoiding the need for complex full-precision multiplication circuits. The distance representation allows for more efficient arithmetic operations compared to traditional binary representations.
4Quantity of substance
If fewer bits are used for quantization, then resource requirements are reduced, but signal accuracy and processing performance deteriorate
Solution Approach 1:
The patent uses dimensional transformation by encoding signal magnitude as distances between non-zero bits rather than direct binary representation. This alternative encoding dimension preserves the relative precision needed for accurate signal processing while using fewer bits, thereby maintaining processing performance and reliability with reduced resource requirements.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
Method, implemented by a processor, of coding a real signal, for example an image signal, into a quantized signal, said method comprising the following steps applied to each real sample of said real signal: • converting (101) the real sample into a numerical representation, • selecting (102), from the fractional part of the number, a predetermined number N of most significant nonzero bits, • for each nonzero significant bit i selected, i varying from 1 to N, determining (103) its distance P, with respect to the neighbouring nonzero significant bit of higher rank selected or, for the first nonzero significant bit selected, with respect to the decimal point, • deducting (104) from said distance P, the minimum value of distance between two nonzero bits, • coding (106) the modified distance P, on a predetermined number M of bits.