Neural Network Parameter Determination for Wireless Signal Distortion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks used in wireless communication systems, such as those for digital pre-distortion, require significant processing resources due to their complex structure, leading to high computational demands.
Innovation Solution
A parameter determination apparatus and method that learns weights and biases for a neural network with a sparse structure by selecting valid connection paths and optimizing weights between layers, reducing the number of connections and processing requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network with complete connection paths is used, then the processing accuracy and distortion compensation performance are improved, but the processing amount and computational complexity increase significantly
Solution Approach 1:
The patent extracts only the necessary connection paths from the complete neural network structure. By identifying and removing redundant connections that do not contribute significantly to the output, the network maintains its distortion compensation accuracy while reducing the number of processing operations required.
Solution Approach 2:
The patent segments the neural network into multiple layers with selective connections. Instead of maintaining complete connections between all layers, the network is divided into segments where only specific connection paths are preserved based on their contribution to the final output, reducing overall complexity.
2Reliability
If a neural network with all possible connection paths is used, then the processing completeness is improved, but the processing amount increases
Solution Approach 1:
The patent applies partial action by maintaining only the necessary connection paths rather than all possible paths. By identifying the minimum set of connections required to achieve reliable processing results, the network maintains processing completeness while improving efficiency by eliminating redundant computations.
3Measurement precision
If the number of connection paths in the neural network is increased, then the learning accuracy is improved, but the calculation amount increases
Solution Approach 1:
The patent extracts and retains only the most significant connection paths that contribute to learning accuracy. By analyzing the contribution of each connection path to the final output, the network removes redundant paths while preserving those essential for accurate learning, thereby reducing computational power requirements.
4Manufacturing precision
If a dense neural network structure is used, then the modeling accuracy is improved, but the processing time increases
Solution Approach 1:
The patent segments the dense network structure into a sparse structure by dividing the connection paths into essential and redundant categories. This segmentation allows the network to maintain modeling accuracy through essential paths while reducing processing time by eliminating redundant computational operations.
Data Source
AI summary
A parameter determination apparatus (2) includes: a first learning device (2111) learning a weight between a [j−1]-th layer (j is an integer that satisfies a condition that “2≤j≤the number of the layer”) and a [j]-th layer to which an output of the [j−1]-th layer is inputted among a plurality of layers of a neural network; a selecting device (2112) selecting at least one valid path for each node included in the [j]-th layer from a plurality of connection paths that connect nodes in the [j−1]-th layer and nodes in the [j]-th layer, respectively, on the basis of the weight learned by the first learning device; and a second learning device (2113) learning at least one of the weight and a bias as the parameters relating to a network structure between the [j−1]-th layer and the [j]-th layer on the basis of the sample signal, the label signal and the valid path.