20240028886. Graph Optimization Method and Apparatus for Neural Network Computation simplified abstract (ZHEJIANG LAB)
Contents
Graph Optimization Method and Apparatus for Neural Network Computation
Organization Name
Inventor(s)
Hongsheng Wang of Hangzhou (CN)
Graph Optimization Method and Apparatus for Neural Network Computation - A simplified explanation of the abstract
This abstract first appeared for US patent application 20240028886 titled 'Graph Optimization Method and Apparatus for Neural Network Computation
Simplified Explanation
The disclosed patent application describes a method and apparatus for optimizing the computation of neural networks. The method includes several steps such as converting a computation graph, allocating a register, defining a route selector for a redefined variable, solving the route selector, defining a criterion for inserting the route selector into a node, analyzing a dominating edge set of the node, inserting the route selector, and renaming the redefined variable.
- The method involves converting a computation graph and allocating a register.
- A route selector is defined for a redefined variable, and the method solves the route selector.
- A criterion is defined for inserting the route selector into a node.
- The method analyzes a dominating edge set of the node for the redefined variable.
- The route selector for the redefined variable is inserted into the node.
- The redefined variable is renamed.
Potential applications of this technology:
- Optimization of neural network computations.
- Improved memory efficiency in deep neural network models.
- Enhanced implementation of neural network models in various applications.
Problems solved by this technology:
- Correct route selection for redefined variables in computation graphs with multiple paths.
- Reducing memory costs in deep neural network models.
Benefits of this technology:
- Improved efficiency and performance of neural network computations.
- Reduction in memory usage and associated costs.
- Facilitates the development and implementation of deep neural network models.
Original Abstract Submitted
the disclosure discloses a graph optimization method and apparatus for neural network computation. the graph optimization method includes the following steps: s: converting a computation graph; s: allocating a register; s: defining a route selector for a redefined variable; s: solving the route selector for the redefined variable; s: defining a criterion of inserting the route selector for the redefined variable into a node; s: analyzing a dominating edge set of the node for the redefined variable; s: inserting the route selector for the redefined variable; and s: renaming the redefined variable. the disclosure solves the problem of the corresponding route selection on a correct definition of the redefined variable when a node including the redefined variable in a computation graph in the compiling period flows through multiple paths of computation flow, reduces the memory cost and promotes the development of implementation application of a deep neural network model.
