Neural Network Frequency Vector Generation via Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network techniques for processing time series data, such as WaveNet, face challenges with high computational costs due to the need for large-scale neural network calculations for each audio sample, and limitations in handling dependencies among dimensions in vector sequence data.
Innovation Solution
A method involving a main neural network and sub-neural networks is employed to generate frequency component vectors of time series data. The main neural network processes initial data to generate intermediate data, which is then used by sub-neural networks to produce component values for different frequency bands, with each sub-neural network's output depending on previous component values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a neural network outputs simultaneous probability distributions of all dimensions or conditional probability distributions based on dependencies among dimensions, then the quality of output is improved, but the number of parameters and calculation resources required become huge
Solution Approach 1:
The patent segments the probability distribution output into separate unidimensional probability distributions for each dimension rather than outputting simultaneous multidimensional probability distributions. This allows the neural network to process each dimension independently, significantly reducing the number of parameters required while maintaining the ability to capture dependencies through sequential processing of unidimensional distributions.
2Manufacturing precision
If large-scale neural network calculation is performed for each audio sample, then the quality of sound sample generation is improved, but the calculation amount becomes huge
Solution Approach 1:
The patent applies segmentation by processing probability distributions unidimensionally rather than simultaneously for all dimensions. This divides the complex high-dimensional probability distribution problem into multiple simpler one-dimensional problems, reducing the calculation amount required for each audio sample while maintaining generation quality through sequential processing of component values.
3Adaptability or versatility
If the number of probability distributions to be outputted by the neural network is increased, then the completeness of dimensional representation is improved, but the calculation resources required for training or generation become huge
Solution Approach 1:
The patent segments the representation of multidimensional probability distributions into separate unidimensional probability distributions. This allows complete representation of all dimensions through independent unidimensional distributions, avoiding the computational explosion that would occur with simultaneous multidimensional probability distributions while maintaining full dimensional coverage.
Data Source
AI summary
A computer-implemented method generates a frequency component vector of time series data, by executing a first process and a second process in each unit step. The first process includes: receiving first data; and processing the first data using a first neural network to generate intermediate data. The second process includes: receiving the generated intermediate data; and generating a plurality of component values corresponding to a plurality of frequency bands based on the generated intermediate data such that: a first component value corresponding to a first frequency band is generated using a second neural network based on the generated intermediate data; and a second component value corresponding to a second frequency band different from the first frequency band is generated using the second neural network based on the generated intermediate data and the generated first component value corresponding to the first frequency band.