Dynamic Neural Network Pruning for Resource-Constrained Hardware
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks in vehicles and embedded systems face resource constraints due to high demands on hardware, especially in safety-critical applications, where existing methods fail to efficiently utilize resources without compromising performance or increasing costs.
Innovation Solution
The method involves network pruning using a negative metric to identify non-essential neurons, followed by selecting a subset of crucial neurons based on a positive metric for specific input data, allowing for dynamic resource allocation and improved output quality within limited hardware capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the full set of neurons in the neural network is used to maintain high processing quality, then the output quality is preserved, but the computing time and hardware resource requirements increase significantly
Solution Approach 1:
The patent segments the full set of neurons into two subsets: P (neurons that can be omitted) and N\P (remaining neurons). This segmentation allows the system to process input data using only the necessary subset of neurons, reducing computing time while maintaining output quality for certain input types.
Solution Approach 2:
The patent dynamically selects which subset of neurons to use based on the input data characteristics. The system determines subset P once, then for each input datum x, it identifies subset A of crucial neurons from N\P and selects D (where D⊇A) for processing. This dynamic adaptation allows optimizing between speed and quality based on actual input requirements.
2Productivity
If more neurons are included in the neural network to improve processing capabilities, then the performance and accuracy increase, but the hardware cost and complexity increase
Solution Approach 1:
The patent extracts and identifies subset P of neurons that can be omitted without unduly impairing performance. By removing these redundant neurons from the active processing path, the system reduces hardware complexity and resource requirements while maintaining the essential processing capabilities needed for accurate output.
Solution Approach 2:
The patent changes the operational parameters of the neural network by selectively activating different subsets of neurons (A and D) based on input data characteristics. This parameter change allows the system to adapt its effective size and complexity dynamically, using only the necessary computational resources for each specific processing task.
3Productivity
If neurons are selectively activated based on positive metric for each input data, then the processing efficiency improves, but the complexity of neuron selection and metric evaluation increases
Solution Approach 1:
The patent performs preliminary action by determining subset P (neurons that can be omitted) once in advance using a negative metric, before processing any specific input data. This pre-computation simplifies subsequent processing by establishing a reduced set N\P from which crucial neurons can be selected, reducing the complexity of per-input evaluation.
Solution Approach 2:
The patent introduces subset A as an intermediary between the pruned network N\P and the final selected subset D. Subset A is determined using a positive metric to identify neurons that significantly contribute to processing specific input data, serving as a bridge that simplifies the selection process while maintaining processing efficiency.
Data Source
AI summary
A method (100) for processing input data x with a neural network (1) that comprises a set N of neurons (11-19), having the steps:according to a given negative metric (2), the subset P⊂N of neurons (11-19) is determined (110) whose use can be omitted without unduly impairing the performance of the neural network (1);according to a given positive metric (3), the subset A⊂(N\P) of those neurons (11-19) that significantly contribute to the processing of the specific input data x is determined (120) from the subset N\P;for processing the specific input data x, a subset D⊂N of neurons (11-19) is selected (130), which is a superset D⊇A of the set A;the input data x are processed into output data y (140) using the neurons (11-19) of the set D.