Hearing Aid Sparse Neural Weights for Low-Memory DNN Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Hearing devices implementing machine learning and deep neural networks (DNNs) face computational costs and efficiency challenges due to their resource-intensive nature.
Innovation Solution
The use of sparse matrix representations for neural network weights in hearing devices, which includes a header indicating the position of non-zero elements, allows for efficient storage and processing, reducing the need to store zeroed submatrices and enabling compact data representation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If dense weight matrices are used for neural network in hearing devices, then computational accuracy is maintained, but memory storage requirements and computational overhead increase significantly
Solution Approach 1:
The patent extracts and removes zero elements from the weight matrices, storing only the non-zero elements along with their position information in header arrays. This extraction principle directly reduces memory storage requirements while maintaining the computational accuracy needed for neural network operations in hearing devices.
Solution Approach 2:
The patent segments the weight matrix into multiple smaller sub-matrices (e.g., 4x4 blocks), where each sub-matrix is processed independently. This segmentation enables efficient sparse matrix operations and reduces the computational overhead by allowing selective processing of only non-zero sub-matrices, while maintaining overall computational accuracy.
2Reliability
If full weight matrices are stored in memory, then complete neural network functionality is preserved, but power consumption and processing efficiency deteriorate
Solution Approach 1:
By extracting and storing only non-zero weight elements, the patent reduces the amount of data that needs to be loaded into memory and processed during neural network operations. This extraction significantly reduces power consumption associated with memory access and computational operations while preserving the essential neural network functionality through the use of header arrays that track non-zero element positions.
3Quantity of substance
If sparse matrix representation is implemented, then storage efficiency is improved, but device complexity increases due to header management
Solution Approach 1:
The patent segments the weight matrix into fixed-size sub-matrices (e.g., 4x4 blocks), which simplifies header management by providing a regular, predictable structure. Each sub-matrix can be independently processed and managed, reducing the overall complexity of header operations while maintaining high storage efficiency through selective storage of only non-zero sub-matrices.
4Ease of operation
If traditional dense matrix operations are used, then computational simplicity is maintained, but processing speed and efficiency decrease
Solution Approach 1:
The patent divides the weight matrix into smaller sub-matrices, enabling parallel processing and optimized computational operations. This segmentation allows the system to process only relevant non-zero sub-matrices, significantly improving processing speed while maintaining computational simplicity through the use of standard matrix multiplication operations on the segmented data structure.
Data Source
AI summary
A hearing device is disclosed. The hearing device comprises a set of input transducers for provision of transducer input data, the set of input transducers comprising a first input transducer for provision of a first transducer input signal as part of the transducer input data. The hearing device comprises a processor for processing transducer input data and providing an electrical output signal based on the transducer input data. The hearing device comprises a receiver for converting the electrical output signal to an audio output signal.