Codebook Feature Vector Output with BMU Distance Updating
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep neural network models trained through global training methods face challenges in achieving high accuracy and efficiency in learning predetermined tasks due to the complexity of updating weight values based on errors between actual and expected outputs.
Innovation Solution
A processor-implemented method that determines distances between input vectors and center vectors in a trained codebook, identifies a best-matched unit (BMU), and updates center vectors based on these distances to improve training accuracy, followed by outputting feature vectors and maps using averages and position coordinates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If global training method with backpropagation is used to train deep neural network models, then training accuracy can be improved, but the complexity of updating weight values and computational requirements increase
Solution Approach 1:
The patent segments the weight updating process into two distinct phases: pre-training phase where global training with backpropagation is performed to establish initial weight values, and post-training phase where only center vectors are updated based on distance calculations. This segmentation reduces the complexity of continuous weight updating while maintaining training accuracy.
Solution Approach 2:
The patent extracts the essential training function into a simplified post-training process that operates on pre-computed distance matrices. By taking out the complex backpropagation updates and replacing them with distance-based center vector adjustments, the system achieves reduced computational complexity while preserving accuracy.
2Productivity
If distance-based center vector updating is applied after pre-training, then computational efficiency improves, but the complexity of determining distances and identifying BMUs increases
Solution Approach 1:
The patent performs preliminary actions by pre-training the neural network to compute distance matrices between input vectors and center vectors before the actual classification task. This pre-computation stores distance information in advance, so that during post-training only simple lookups and updates are needed, significantly reducing real-time computational complexity.
Solution Approach 2:
The patent creates a simplified copy of the training process for post-training purposes. Instead of performing full backpropagation, the system uses pre-computed distance information to update center vectors, effectively copying the essential training function into a more efficient form that maintains accuracy while reducing computational burden.
3Measurement precision
If feature vectors and maps are output based on distance calculations, then data processing precision improves, but the complexity of processing and outputting operations increases
Solution Approach 1:
The patent implements self-service by having the system automatically determine distances, identify best-matched units, and update center vectors without requiring complex external processing. The distance-based mechanism inherently provides precise feature vector output through straightforward calculations, reducing the need for additional complex processing operations.
Data Source
AI summary
A processor-implemented method includes: determining distances between an input vector and center vectors comprised in a plurality of output nodes comprised in a trained codebook; and outputting a first feature vector of the input vector based on the distances between the center vectors and the input vector, wherein the trained codebook is trained by: determining a distance between a training input vector and the center vector for each of the output nodes; determining, among the plurality of output nodes, a best matched unit (BMU) in which a distance between the training input vector and the center vector of the BMU is minimized; and training the codebook by updating the center vector of the BMU, based on the distance between the training input vector and the center vector of the BMU.