Neural Network Index-Based Processing for Power Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network devices face challenges in efficiently processing complex input data in real-time due to high power consumption and operational inefficiencies, particularly in performing neural network operations such as convolution and pooling.
Innovation Solution
The implementation of an index-based neural network operation method, where an input feature list is generated including input feature indices and values, with operations performed based on these indices to reduce redundant calculations, specifically by adding input feature indices and weight indices, dividing by an integer, and selecting the quotient as an output feature index, thereby skipping operations on zero-value inputs and weights.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional neural network operations are performed on all input data, then complete processing is achieved, but power consumption increases and processing speed decreases
Solution Approach 1:
The patent extracts and processes only the meaningful non-zero elements from the input feature map and weight map, storing them in separate lists with their corresponding indices. This extraction principle eliminates redundant zero-value operations, directly reducing power consumption and improving processing speed by focusing computational resources only on relevant data elements.
Solution Approach 2:
Instead of performing complete operations on all elements of the input and weight maps, the patent applies partial action by selectively processing only the non-zero elements. The index-based addressing mechanism enables the system to perform exactly the necessary computations without excessive operations on zero-value elements, optimizing both energy efficiency and processing speed.
2Productivity
If operations are performed on all elements including zero-values, then no data is lost, but the number of operations increases reducing efficiency
Solution Approach 1:
The patent extracts only the non-zero elements from the input feature map and weight map into separate lists, accompanied by their index information. This extraction eliminates the need to process zero-value elements, directly reducing the number of operations and processing time while maintaining operational efficiency by focusing only on meaningful computations.
Solution Approach 2:
The patent performs preliminary action by pre-processing the input feature map and weight map to identify and store non-zero elements with their indices before the main neural network operations. This preliminary extraction step enables subsequent operations to proceed more efficiently by avoiding redundant computations on zero-values, thereby reducing overall processing time.
3Productivity
If index-based operations with division and quotient selection are implemented, then redundant operations are reduced, but computational complexity increases
Solution Approach 1:
The patent introduces an intermediary mechanism using index lists that store the positions of non-zero elements. This intermediary structure enables the system to efficiently locate and process only relevant elements without performing division and quotient operations on all elements. The index lists act as a mediator between the input data and processing operations, reducing the overall computational complexity while maintaining the benefit of reduced redundant operations.
Data Source
AI summary
A neural network device may generate an input feature list based on an input feature map, where the input feature list includes an input feature index and an input feature value, generating an output feature index based on the input feature index corresponding to an input feature included in the input feature list and a weight index corresponding to a weight included in a weight list, and generating an output feature value corresponding to the output feature index based on the input feature value corresponding to the input feature and a weight value corresponding to the weight.