Keyword-Spotting CNN Feature Reuse Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current keyword-spotting neural network circuits face high power consumption due to frequent access to storage modules for parallel calculations, leading to inefficiencies in energy usage and resource occupancy.
Innovation Solution
A speech feature reuse-based storing and calculating compression method is implemented, where only the newest frame of input data is stored, and intermediate data is minimized to the size of the convolution kernel, reducing redundant calculations and storage, and utilizing a frame reuse calculation mode to optimize convolution operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If parallel calculation is performed in the neural network processor, then the calculation speed is improved, but the access storage frequency increases and power consumption increases
Solution Approach 1:
The patent performs preliminary actions by pre-calculating and storing convolution results for multiple frames in advance. The storage module keeps historical convolution results (e.g., from frames t-1, t-2, etc.) ready for reuse, so that when new frames arrive, the system can directly retrieve and combine these pre-computed results without re-calculating, thus reducing real-time calculation load and power consumption while maintaining high processing speed
Solution Approach 2:
The patent uses copying by creating and storing copies of convolution results from previous frames. Instead of recalculating convolution operations for each new frame, the system copies and reuses the computational results from historical frames, combining them with new input data through addition operations. This copying approach dramatically reduces the calculation burden and associated power consumption while preserving the parallel processing capability's speed benefits
2Productivity
If parameters and intermediate data are dramatically increased for parallel calculation, then the calculation capability is improved, but the storage module access frequency increases and resource occupancy increases
Solution Approach 1:
The patent applies copying by storing copies of intermediate convolution results in the storage module. These copied results from previous frames are retained and reused for subsequent calculations, eliminating the need to maintain large amounts of intermediate data for every possible calculation scenario. This reduces the storage requirements and resource occupancy while preserving high calculation capability through efficient data reuse
Solution Approach 2:
The patent implements discarding and recovering by selectively discarding old convolution results when new frames arrive, and recovering (retrieving) only the specific historical results needed for the current calculation. The storage module manages a limited set of recent convolution results, discarding older ones as newer frames arrive, and recovering only the necessary historical data for combining with current inputs. This approach maintains calculation capability while significantly reducing resource occupancy compared to storing all possible intermediate data
Data Source
AI summary
It discloses a speech feature reuse-based storing and calculating compression method for a keyword-spotting CNN, and belongs to the technical filed of calculating, reckoning or counting. If the updated row number of input data is equal to a convolution step size, every time new input data arrive, an input layer of a neural network replaces the earliest part of the input data with the new input data and meanwhile adjusts an addressing sequence of the input data, thereby performing an operation on the input data and corresponding convolution kernels in an arrival sequence of the input data, and an operation result is stored in an intermediate data memory of the neural network to update corresponding data.