Neural Network Distortion Compensation Circuit Size Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Signal transmission apparatuses face challenges in reducing circuit size due to the need for large memory capacity to store numerous parameter sets for Neural Networks used in distortion compensation, as the number of parameters is enormous and varies with signal patterns, leading to increased circuit size.
Innovation Solution
A signal transmission apparatus with a Neural Network having L+1 arithmetic layers, including a hidden layer and an output layer, where only the output layer's parameters are switched based on signal patterns, and other layers' parameters are fixed, reducing the number of parameter sets stored and thus the circuit size.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all parameters of the Neural Network are stored for each signal pattern, then distortion compensation accuracy is improved, but memory capacity and circuit size increase enormously
Solution Approach 1:
The Neural Network parameters are segmented into two groups: common parameters shared across all signal patterns and signal-specific parameters unique to each pattern. This segmentation allows the system to store only the necessary signal-specific parameters in memory, dramatically reducing storage requirements while maintaining compensation accuracy.
Solution Approach 2:
The common parameters serve multiple signal patterns simultaneously, making them universal across different input conditions. By identifying and extracting these universal parameters, the system avoids redundant storage and achieves efficient memory utilization without sacrificing distortion compensation performance.
2Measurement precision
If all parameters of the Neural Network are stored for each signal pattern, then distortion compensation accuracy is improved, but circuit size increases
Solution Approach 1:
The parameter storage structure is segmented into common parameter storage and signal-specific parameter storage. This segmentation simplifies the circuit architecture by eliminating the need for enormous dedicated storage for each signal pattern, thereby reducing overall circuit size while preserving accuracy.
Solution Approach 2:
Common parameters are extracted from the full parameter sets and stored separately. This extraction reduces the amount of data that needs to be stored in signal-specific memory structures, directly reducing circuit complexity and size.
3Quantity of substance
If the number of parameter sets is reduced by fixing hidden layer parameters, then memory capacity is reduced, but distortion compensation adaptability may be compromised
Solution Approach 1:
Different parts of the Neural Network are assigned different qualities: hidden layer parameters are fixed to provide stability and reduce memory requirements, while output layer parameters are made signal-specific to maintain adaptability. This local differentiation of parameter characteristics resolves the contradiction between memory reduction and adaptability preservation.
Solution Approach 2:
The system dynamically changes which parameters are fixed and which are variable based on signal patterns. By fixing hidden layer parameters and allowing output layer parameters to vary with signal patterns, the system achieves both memory efficiency and compensation adaptability.
Data Source
AI summary
A signal transmission apparatus (1) includes: a distortion compensation unit (11) for performing a distortion compensation processing on an input signal (x) by using a Neural Network (112) including L+1 arithmetic layers that include L (L is a variable number representing an integer equal to or larger than 1) hidden layer (112M) and an output layer (112O); a storage unit (13) for storing parameter sets (131) each of which includes a parameter for Q (Q is a variable number representing an integer equal to or smaller than L) arithmetic layer of the L+1 arithmetic layers; and an application unit (142) for selecting one parameter set from the parameter sets based on a signal pattern of the input signal and applying the parameter included in the selected one parameter set to the M number of arithmetic layer, a parameter of another arithmetic layer of the L+1 arithmetic layers, which is other than the Q arithmetic layer, is fixed.