Neural Network Parameter Coding with Inferred Syntax and History Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for encoding and decoding neural network parameters, such as deepCABAC, face challenges in memory requirements and bitrate due to the need to encode/decode all syntax elements, and inefficiently handle updates where most parameters remain unchanged.
Innovation Solution
The proposed solution involves inferring predetermined syntax elements based on a mapping scheme used during quantization/dequantization, and selecting a probability model for update parameters based on their current and previous values to optimize encoding/decoding efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all syntax elements are encoded/decoded for neural network parameters, then complete parameter representation is achieved, but memory requirements and bitrate increase
Solution Approach 1:
The patent extracts and removes redundant syntax elements from the encoding/decoding process. By identifying that certain syntax elements can be inferred from mapping schemes or remain unchanged during updates, the invention selectively excludes these elements from transmission, thereby reducing memory requirements and bitrate while preserving essential parameter information.
Solution Approach 2:
Instead of encoding/decoding all syntax elements completely, the patent applies partial action by only encoding/decoding necessary elements. For update parameters, only changed values are encoded with selected probability models, and for non-updated parameters, inference is used instead of full encoding, achieving sufficient representation with reduced data volume.
2Reliability
If all syntax elements are encoded/decoded for neural network parameters, then complete parameter representation is achieved, but bitrate increases
Solution Approach 1:
The patent extracts and removes redundant syntax elements from the encoding/decoding process. By identifying that certain syntax elements can be inferred from mapping schemes or remain unchanged during updates, the invention selectively excludes these elements from transmission, thereby reducing memory requirements and bitrate while preserving essential parameter information.
Solution Approach 2:
Instead of encoding/decoding all syntax elements completely, the patent applies partial action by only encoding/decoding necessary elements. For update parameters, only changed values are encoded with selected probability models, and for non-updated parameters, inference is used instead of full encoding, achieving sufficient representation with reduced data volume.
3Ease of manufacture
If traditional entropy coding is used for update parameters, then encoding is simple, but coding efficiency decreases
Solution Approach 1:
The patent introduces dynamic probability model selection for entropy coding update parameters. Instead of using a fixed coding scheme, the invention adapts the probability model based on the current and previous values of parameters, allowing the coding efficiency to dynamically adjust to the actual data characteristics while maintaining implementation feasibility.
Solution Approach 2:
The patent changes the parameters of the entropy coding process by selecting different probability models based on parameter values. This involves monitoring parameter changes and adjusting the coding strategy accordingly, transforming a static coding process into a adaptive one that optimizes efficiency based on actual parameter behavior.
Data Source
AI summary
Some embodiments relate to a method, a decoder and/or an encoder for entropy coding of parameters of neural networks and their incremental updates, and in particular to reduced value set coding and history depended significance coding.