Ternary Weight Compression Using 4-Bit Flags and 3-Bit Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for compressing ternary weight data in neural networks cannot reduce the bit count further than existing techniques, leading to memory and bandwidth inefficiencies.
Innovation Solution
A method that generates a 4-bit data string by dividing ternary weight data into 4-bit items, using a flag value and non-zero value strings to represent the data, where the flag value is assigned to a 4-bit item '0000' and other items are converted into 3-bit values, reducing the bit count by utilizing high appearance frequencies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional compression methods (ZVC or RLE) are used on ternary weight data, then the bit count is reduced to some extent, but the bit count cannot be reduced any further
Solution Approach 1:
The patent segments the ternary weight data into 4-bit data items and processes them in groups. By dividing the data into manageable segments and applying flag-based compression at the 4-bit level rather than the traditional 2-bit level, the method achieves finer-grained compression control and higher overall compression ratios.
Solution Approach 2:
The patent changes the fundamental parameter of data representation by using 4-bit items instead of traditional 2-bit items, and by introducing a flag value system where one value (0000) is represented by a single bit flag while other values are encoded differently. This parameter change enables achieving 50% or more compression ratios.
2Adaptability or versatility
If high-dimensional weight data is used in convolutional operation, then the neural network can process complex data, but the memory region and bus bandwidth are strained
Solution Approach 1:
The patent extracts and eliminates redundant information from the weight data by identifying that ternary values (0, 1, -1) can be efficiently represented through a flag system. By taking out only the essential information needed to reconstruct the original data and discarding redundant representations, significant memory and bandwidth savings are achieved.
Solution Approach 2:
The patent transitions from traditional 2-bit representation to a 4-bit flag-based representation system, effectively changing the dimensional approach to data encoding. This dimensional change allows for more efficient packing of information and achieves higher compression ratios while maintaining the ability to represent all necessary ternary weight values.
Data Source
AI summary
A weight data compression method includes: generating a 4-bit data string of 4-bit data items each expressed as any one of nine 4-bit values, by dividing ternary weight data into data items each having 4 bits; and generating first compressed data including a first flag value string and a first non-zero value string by (i) generating the first flag value string by assigning one of 0 and 1 as a first flag value of a 1-bit flag to a 4-bit data item 0000 and assigning an other of 0 and 1 as a second flag value of the 1-bit flag to a 4-bit data item other than 0000 among the 4-bit data items in the 4-bit data string and (ii) generating the first non-zero value string by converting the 4-bit data item other than 0000 into a 3-bit data item having any one of eight 3-bit values.