Grouped Feature Map Quantization for Neural Network Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current video compression technologies face challenges in efficiently encoding and decoding feature maps from convolutional neural networks, particularly in managing quantization parameters across multiple feature maps, which affects processing efficiency and flexibility in distributed systems.
Innovation Solution
A method and apparatus for determining and applying shared quantization parameters to decode or encode feature maps, allowing for efficient inverse quantization or quantization across groups of feature maps, thereby optimizing processing and reducing overhead in distributed systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate quantization parameters are used for each feature map, then quantization precision can be optimized for each map, but processing complexity and computational burden increase significantly
Solution Approach 1:
The patent merges quantization parameters across multiple feature maps by introducing a shared quantization parameter that is common to all feature maps in a given layer. This consolidation reduces the number of separate quantization operations from O(N) to O(1), significantly lowering processing complexity while maintaining acceptable quantization precision through the shared parameter's ability to uniformly control all feature maps.
Solution Approach 2:
The shared quantization parameter serves multiple feature maps simultaneously, making it a universal parameter that performs the quantization function for all maps in a layer. This multi-functionality eliminates the need for individual quantization parameters for each feature map, reducing computational burden while maintaining system functionality.
2Manufacturing precision
If individual quantization parameters are assigned to each feature map, then encoding accuracy is improved, but data transmission overhead increases
Solution Approach 1:
The patent combines multiple individual quantization parameters into a single shared quantization parameter that is transmitted once for all feature maps. This reduces the amount of quantization parameter data that needs to be transmitted through the bitstream, significantly lowering data transmission overhead. The shared parameter maintains encoding accuracy by uniformly controlling quantization across all feature maps in the layer.
Solution Approach 2:
The patent changes the parameter structure from multiple individual quantization parameters to a single shared quantization parameter. This parameter change reduces the quantity of parameters that need to be encoded and transmitted, thereby reducing overhead while maintaining the functional capability of precise quantization control across all feature maps.
3Productivity
If quantization parameters are shared across feature maps, then processing efficiency improves, but quantization flexibility is reduced
Solution Approach 1:
The patent segments feature maps into groups (such as by layer or by computational graph structure) and applies a single shared quantization parameter to each segment. This segmentation approach maintains processing efficiency by using a single parameter per segment while preserving quantization flexibility at the segment level, allowing different segments to have different shared parameters tailored to their specific characteristics.
Solution Approach 2:
The patent applies local quality by allowing different shared quantization parameters for different segments or groups of feature maps. Each segment can have its own shared parameter optimized for its specific characteristics, maintaining flexibility at the local level while achieving efficiency through sharing within each segment. This balances processing efficiency with quantization flexibility by enabling segment-specific optimization.
Data Source
AI summary
A method of decoding feature maps from encoded data. A parameter related to quantisation is determined. According to the parameter, inverse quantisation is performed for samples decoded from the encoded data to decode the feature maps. For at least a part of the samples decoded from the encoded data, one parameter is shared by a group of samples corresponding to a plurality of feature maps.