Non-Maximum Suppression Filtering With Statistical Score Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The computational time complexity of non-maximum suppression (NMS) operations in neural networks is high due to sorting operations, and transferring large datasets to host memory for processing introduces significant latency, especially for large input sizes.
Innovation Solution
Perform a statistical analysis to discard low-scoring candidate predictions using mean and standard deviation calculations, allowing initial search space reduction within the neural network accelerator, reducing the number of candidates transferred to host memory.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional sorting-based NMS operation is used, then prediction accuracy is maintained, but computational time complexity increases significantly
Solution Approach 1:
The patent changes the approach from sorting-based to statistical analysis-based filtering. Instead of sorting all candidate predictions by score and selectively keeping top candidates, the system uses mean and standard deviation calculations to determine a threshold and filters candidates accordingly. This parameter change in the filtering methodology reduces computational complexity from O(N log N) sorting to O(N) statistical computation while maintaining prediction accuracy.
Solution Approach 2:
The patent extracts the essential function of NMS (filtering out low-quality predictions) while removing the computationally expensive sorting operation. By taking out the sorting step and replacing it with statistical analysis, the system maintains the core functionality of eliminating redundant predictions without the associated computational burden.
2Productivity
If large candidate predictions are transferred to host memory for processing, then processing capability is utilized, but memory latency increases significantly
Solution Approach 1:
The patent performs preliminary filtering of candidate predictions within the neural network accelerator using statistical analysis before transferring data to host memory. By pre-computing mean and standard deviation and applying threshold filtering in advance, the system reduces the volume of data that needs to be transferred, thereby reducing memory latency and improving overall processing efficiency.
Solution Approach 2:
The patent segments the NMS operation into two parts: a filtering stage performed within the accelerator using statistical analysis, and a transfer stage to host memory for final processing. This segmentation allows the computationally intensive filtering to be done locally without requiring large data transfers, reducing memory latency.
3Loss of time
If statistical analysis is performed for search space reduction, then processing time is reduced, but computational operations increase
Solution Approach 1:
The patent changes the computational approach from complex sorting operations to simpler statistical analysis operations. By using mean and standard deviation calculations followed by threshold comparison, the system achieves O(N) time complexity, which is faster than sorting-based O(N log N) approaches, while the computational operations remain relatively simple arithmetic operations.
Data Source
AI summary
A technique for improving the computational time for performing a non-maximum suppression operation may include receiving a request to perform a non-maximum suppression operation on a set of candidate predictions of a computing task, and performing a statistical analysis on a set of confidence scores to reduce the set of candidate predictions. Candidate predictions from the reduced set of candidate predictions are provided as the result of the non-maximum suppression operation.