Variable-Length Feature Extraction with Multiscale Maps and Recursive Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for processing variable-length data require a minimum length for feature maps, limiting their applicability to data with varying lengths.
Innovation Solution
An information processing apparatus that generates scale-specific feature maps, series, and information by using convolutional layers and recursive models to handle data of varying lengths.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiscale statistics pooling is used to extract features from variable-length speech, then feature extraction capability is improved, but the technique requires a minimum data length which limits applicability to diverse length data
Solution Approach 1:
The patent segments the input variable-length data into multiple fixed-length segments, allowing the model to process data of any length by dividing it into manageable pieces. This resolves the contradiction by enabling the fixed-structure pooling operation to work on segmented data rather than requiring the entire input to meet a minimum length requirement.
Solution Approach 2:
The patent introduces a segmentation dimension by transforming the problem from processing variable-length sequences directly to processing fixed-length segments. This dimensional transformation allows the model to handle diverse length data by operating in the segment space rather than the original variable-length space.
2Productivity
If fixed-structure pooling operations are used, then computational efficiency is improved, but the model cannot handle variable-length inputs effectively
Solution Approach 1:
By segmenting variable-length inputs into fixed-length pieces, the patent enables the use of efficient fixed-structure pooling operations on each segment independently, while still maintaining the ability to handle variable-length overall inputs through the combination of multiple segments.
Solution Approach 2:
The patent introduces dynamic segmentation where the number of segments depends on the input length, allowing the model to adapt its processing structure dynamically while maintaining fixed-structure operations within each segment for computational efficiency.
Data Source
AI summary
Provided is at least one processor included in an information processing apparatus, carrying out: a feature map generation process of generating a plurality of scale-specific feature maps from input data; a feature series generation process of generating a feature series from the plurality of scale-specific feature maps; and a feature information generation process of generating feature information by inputting the feature series into a recursive model.