Hearing Device Neural Processing With Fixed-Point Weights
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Hearing devices implementing deep neural networks (DNNs) face computational inefficiencies and potential performance degradation due to the high computational costs and data type mismatches between transducer input data and neural network weights.
Innovation Solution
A hearing device that processes transducer input data using a combination of floating point numbers for improved precision and fixed point numbers for reduced computational cost, allowing efficient conversion and processing without data type conversion, thereby enabling simpler multipliers and reduced error buildup.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating point numbers are used for neural network processing, then precision and dynamic range are improved, but computational cost and power consumption increase
Solution Approach 1:
The patent applies different data types to different components of the neural network processing pipeline. Floating point numbers are used for transducer input data where high precision is critical, while fixed point numbers are used for neural network weights and internal computations where computational efficiency is prioritized. This local differentiation resolves the contradiction by optimizing each component for its specific requirements rather than using a uniform data type throughout.
Solution Approach 2:
The patent changes the data type parameter from uniform floating point to a hybrid system where fixed point numbers are used for neural network operations. This parameter change reduces computational complexity and power consumption while maintaining sufficient precision through the fixed point representation, thereby resolving the trade-off between precision and energy efficiency.
2Adaptability or versatility
If data type conversion is performed between transducer input data and neural network weights, then compatibility is achieved, but processing efficiency decreases
Solution Approach 1:
The patent performs data type alignment as a preliminary action by storing neural network weights in the same fixed point data type as the internal computations. This pre-prepared state eliminates the need for runtime data type conversion between transducer input data and weights, maintaining compatibility while maximizing processing efficiency through direct computation.
Solution Approach 2:
The patent introduces fixed point numbers as an intermediary representation layer between the floating point transducer input data and the neural network computations. This intermediary allows the system to maintain compatibility with high-precision input data while enabling efficient fixed point arithmetic for the neural network operations, thereby resolving the efficiency-compatibility trade-off.
3Adaptability or versatility
If complex data type conversions are implemented, then data compatibility is improved, but device complexity increases
Solution Approach 1:
The patent segments the data processing pipeline into distinct stages with specific data type requirements. Transducer input data maintains floating point precision for accuracy, while neural network weights and internal representations use fixed point for efficiency. This segmentation allows each stage to operate optimally without requiring complex conversion mechanisms, thereby reducing overall device complexity while maintaining data compatibility.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A hearing device and related method is disclosed, the hearing device comprising 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; a processor for processing transducer input data and providing an electrical output signal based on the transducer input data; and a receiver for converting the electrical output signal to an audio output signal, wherein the processor is configured to apply a neural network to a network input based on the transducer input data for provision of a network output, the electrical output signal based on the network output, wherein the network input has a first data type and weights of the neural network have a second data type different from the first data type.