Multi-scale RBF Neural Network Blocking Layer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional Radial Basis Function (RBF) neural networks face challenges in preventing successfully mapped neighborhoods from later node influence, particularly when dealing with overlapping receptive fields, which affects their ability to accurately model and distinguish signals of different resolutions.
Innovation Solution
A multi-scale radial basis function neural network architecture is introduced, incorporating a blocking layer and an algorithm that utilizes local statistics in addition to global Mean Square Error (MSE) to selectively block successfully mapped signals, allowing for the exclusion of later node influence and improved signal isolation in overlapping receptive fields.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional RBF neural networks use overlapping receptive fields to model signals, then the network can capture local patterns, but successfully mapped neighborhoods cannot be excluded from later node influence causing interference
Solution Approach 1:
The patent introduces a blocking layer that segments the network into distinct functional regions. Each blocking function creates a mask that prevents successfully mapped neighborhoods from influencing subsequent nodes, effectively partitioning the receptive field coverage while maintaining the benefits of overlapping fields for pattern capture
Solution Approach 2:
The blocking layer acts as an intermediary between the activation layer and output layer. It receives activation signals, applies blocking masks based on successfully mapped neighborhoods, and passes only unblocked signals forward, thereby mediating the interference problem while preserving useful signal information
2Reliability
If the neural network processes all signals from activation functions, then comprehensive signal coverage is achieved, but computation time increases and convergence slows
Solution Approach 1:
The blocking layer extracts and removes successfully mapped neighborhood signals from the processing stream. By identifying and taking out these already-processed signals, the network avoids redundant computation while maintaining complete signal coverage for unprocessed regions
Solution Approach 2:
Instead of processing all signals equally, the blocking layer applies partial action by selectively blocking only the successfully mapped portions. This reduces the excessive computation on redundant signals while maintaining necessary processing coverage, thereby accelerating convergence without sacrificing reliability
3Device complexity
If the network uses single-scale activation functions, then the architecture is simpler, but it cannot distinguish signals at different resolutions in overlapping fields
Solution Approach 1:
The patent employs multi-scale activation functions with dynamically adjustable receptive field sizes. Different activation functions can adapt their scale to match the local characteristics of the data, enabling the network to distinguish signals at different resolutions while maintaining architectural flexibility
Solution Approach 2:
Different regions of the input space are assigned different activation function scales based on local data characteristics. This local quality approach allows the network to use finer-grained activation functions in regions requiring high resolution and coarser functions in smoother regions, optimizing both precision and complexity
Data Source
AI summary
A network architecture of radial basis function neural network system utilizes a blocking layer (4) to exclude successfully mapped neighborhoods from later node influence. A signal is inserted into the system at input nodes (I1, I2, . . . In), which then promulgates to a non-linear layer (2). The non-linear layer (2) comprises a number of non-linear activation function nodes (10). After passing through the non-linear layer (2), the signal passes through the blocking layer (4) that is comprised of either binary signal blocking nodes, or inverted symmetrical Sigmoidal signal blocking nodes (12) that act in a binary fashion. Finally, the signal is weighted by a weighting function (6a, 6b, 6c, 6n), summed at a summer (8) and outputted at (O).