Custom Operator Decouples Candidate Data From Neural Network Graph
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning graph computing frameworks face limitations in data storage and update freedom, leading to computational overhead and difficulty in parameter transfer and maintenance, especially in big data scenarios like recommendation and search services.
Innovation Solution
A data service method and device that utilize a custom operator within the neural network graph structure to acquire and update candidate datasets outside the graph structure, decoupling them from the graph structure and reducing computational overhead through a hybrid graph computing framework.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If candidate data is solidified into the neural network graph structure in advance (usage mode inside the graph), then the data can be efficiently serialized and distributed, but the freedom of storing and updating candidate data is limited and the data is bound to the graph structure
Solution Approach 1:
The patent segments the data processing system into two independent parts: the graph computing framework (neural network structure) and the candidate data management system. By separating data acquisition, processing, and graph computation into distinct modules, the system allows candidate data to be updated independently without affecting the graph structure, while still enabling efficient serialized processing when needed.
Solution Approach 2:
The patent implements dynamic data loading where candidate data can be flexibly acquired from external sources during graph computation rather than being statically embedded. The system dynamically determines whether to load data from external storage or use pre-serialized data, allowing adaptive updates without retraining the graph structure.
2Adaptability or versatility
If candidate data is used outside the graph structure (usage mode outside the graph), then the freedom of data storage is improved, but it causes extra computational overhead for parameter transfer and maintenance
Solution Approach 1:
The patent merges the benefits of both usage modes by integrating external data access capabilities directly into the graph computing framework through custom operators. This allows the system to access externally stored candidate data without separate transfer steps, eliminating redundant computational overhead while maintaining data flexibility.
Solution Approach 2:
The patent introduces custom operators as intermediaries between the external data storage system and the graph computing framework. These operators handle data transfer and format conversion efficiently, reducing the computational overhead associated with parameter transfer while maintaining the flexibility of external data management.
3Adaptability or versatility
If candidate data is prepared outside the deep learning graph computing framework, then the freedom of data storage is improved, but the difficulty of transferring and maintaining parameters in the graph structure increases
Solution Approach 1:
The patent creates a universal data interface layer that works with both internal and external data sources. The custom operators provide standardized methods for data acquisition, processing, and integration that work consistently across different storage locations, simplifying parameter maintenance regardless of whether data comes from internal or external sources.
Data Source
AI summary
The present application discloses a data service method and device, and related products. After a process of a target data service is started, a candidate dataset outside a neural network graph structure and corresponding to a target data service is obtained through a custom operator in the neural network graph structure; the neural network graph structure and the candidate dataset obtained by the custom operator in the previous step are then used to execute the target data service. A hybrid graph computing framework that combines a usage mode inside the graph with a usage mode outside the graph is realized with the aid of the custom operator. The custom operator is located inside the graph, and the candidate data is located outside the graph. The custom operator can access the candidate data outside the graph, so that the candidate data can participate in the target data service.