Neural Network Coefficient Transmission via Custom Binary Floating Point

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Neural networks face inefficiencies in transmitting large lists of numerical coefficients due to high data volume and the difficulty in compressing numerical data with traditional algorithms, which affects data rate and accuracy.

Innovation Solution

The method involves converting coefficients to a binary floating-point format, determining an exponent prefix and fraction suffix based on the coefficient range and precision, and transmitting only a sign bit, exponent suffix, and fraction prefix, with a header containing the exponent prefix and fraction suffix, allowing for efficient compression and decompression.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If traditional compression algorithms are used to compress coefficient lists, then compression ratio is improved, but measurement precision deteriorates because numerical data has high entropy and is difficult to compress

Engineering Contradiction:
Improvedata sizeVSAvoidcoefficient accuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent changes the representation parameters of coefficients by converting them to floating-point format and then to a customized binary format with variable precision. The system determines the minimum number of bits required to represent each coefficient accurately by examining the coefficient list, then transmits only those necessary bits along with range information, achieving both compression and precision maintenance.

Inventive Principle:
Principle #35Parameter changes

2Productivity

If the number of bits transmitted per coefficient is reduced, then data rate is improved, but measurement precision deteriorates

Engineering Contradiction:
Improvedata transmission rateVSAvoidcoefficient accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent applies local quality by transmitting different precision levels for different coefficients based on their actual requirements. Instead of uniform precision, the system examines each coefficient's magnitude and determines the minimum bits needed to represent it accurately. Coefficients with smaller magnitudes require fewer bits than those with larger magnitudes, optimizing the balance between transmission speed and precision.

Inventive Principle:
Principle #3Local quality

3Device complexity

If coefficients are transmitted in fixed-point format, then device complexity is reduced, but data size increases and compression becomes difficult

Engineering Contradiction:
Improveprocessing complexityVSAvoiddata size
Core Design Contradiction:
Device complexityVSQuantity of substance

Solution Approach 1:

The patent transforms coefficients from fixed-point representation to floating-point representation, then to a customized binary format. This parameter change allows the system to represent coefficients with variable precision using fewer bits. The floating-point format naturally handles different magnitudes of coefficients, enabling efficient compression without significantly increasing processing complexity.

Inventive Principle:
Principle #35Parameter changes

4Measurement precision

If the exponent prefix and fraction suffix are transmitted for each coefficient, then measurement precision is maintained, but data size increases

Engineering Contradiction:
Improvecoefficient accuracyVSAvoiddata size
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent merges the exponent prefix and fraction suffix into a single transmitted unit for each coefficient. Instead of transmitting these components separately or for all coefficients, the system combines them into a compact binary representation that can be efficiently processed. This merging reduces the overall data size while maintaining the necessary precision information for accurate coefficient reconstruction.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentUS10528641B2Encoder and decoder for transmission of coefficients to a neural network
Publication Date: 2020.01.07 CEREMORPHIC INC
  • US10528641B2 patent drawing
  • US10528641B2 patent drawing
  • US10528641B2 patent drawing

AI summary

A method for efficient transmission of coefficients examines a coefficient list, presents the coefficients as binary floating point representation, and transmits the list of coefficients as a header having an exponent prefix, a fractional suffix, and each coefficient value as an exponent suffix and fractional prefix. A method for reception of coefficients receives a header including an exponent prefix, a fractional suffix, thereafter receiving each value as a sign bit, an exponent suffix and a fractional prefix, reconstituting an approximation of the original value, in sequence, as a sign bit, exponent prefix exponent suffix, fraction prefix, and fraction suffix, thereby greatly reducing the amount of information to be transmitted or received.