Tensor Compression With Dynamic Weight Switching for Edge CNNs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) require high computational complexity, which is beyond the capacity of edge devices, necessitating distributed processing across edge devices and cloud servers, but existing compression methods lack adaptability to varying input data and result in sub-optimal performance.
Innovation Solution
A method for encoding and decoding tensors using a hierarchical representation, involving a bottleneck encoder and decoder with trainable weights, allowing for live refinement training to adapt to changing input statistics and maintain optimal performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If distributed processing is used to reduce computational complexity on edge devices, then processing capability is improved, but compression performance deteriorates due to lack of adaptability
Solution Approach 1:
The patent implements dynamic weight updates in the compression and decompression neural networks. The weights are initially trained offline but can be updated online based on feedback from the decompression performance, allowing the system to adapt to varying input data statistics and maintain optimal compression performance in distributed processing environments.
Solution Approach 2:
The patent establishes a feedback mechanism where the decompression output is evaluated and used to update the compression network weights. This closed-loop system allows the compression performance to be continuously improved based on actual decompression results, resolving the performance deterioration issue in distributed processing.
2Device complexity
If fixed weights are used in compression methods, then device complexity is reduced, but adaptability to varying input data deteriorates
Solution Approach 1:
The patent transitions from static fixed weights to dynamic weights that can be updated online. The compression network maintains simple fixed weights during normal operation but can receive weight updates from a training server based on actual input data statistics, achieving both low device complexity and high adaptability.
Solution Approach 2:
The patent performs preliminary weight training offline on powerful servers before deployment to edge devices. This preliminary action prepares the weights in advance, and the system can later refine these weights online based on actual usage, combining the benefits of pre-processing and real-time adaptation.
3Ease of manufacture
If traditional compression methods are used, then implementation simplicity is maintained, but performance optimality deteriorates
Solution Approach 1:
The patent uses neural network copies of the compression and decompression functions that can be deployed to edge devices. These neural network copies maintain the simplicity of traditional compression methods in terms of deployment while achieving superior performance through learned optimal compression strategies.
Solution Approach 2:
The patent changes the parameters of the compression system from fixed traditional compression settings to learnable neural network parameters. These parameters are initially set through training but can be dynamically adjusted based on input data characteristics, achieving performance optimality while maintaining implementation simplicity through automated parameter management.
Data Source
AI summary
An apparatus and method for encoding. The method comprises performing data compression for data related to an image, the data compression using a neural network associated with a first set of weights; determining whether a set of weights for the data compression is to be changed; and encoding first information indicating whether the set of weights is to be changed, based on the determination. If it is determined that the set of weights for the data compression is to be changed, the method comprises changing association of the neural network used in the data compression to a second set of weights from the first set of weights; encoding second information indicating the second set of weights; and encoding the data in which the data compression is performed using the neural network with the associated second set of weights.