Neural Network API Using Descriptor-Based Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning libraries often fail to provide comprehensive optimizations for parallel processing units and have complex application programming interfaces (APIs), leading to decreased maintainability and performance for software programs utilizing these libraries.
Innovation Solution
A deep neural network library architecture that facilitates optimized neural network computations by using descriptors provided by users to generate optimized operation sets, select computational options, and manage data pointers, thereby improving performance and maintainability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If comprehensive optimizations are provided for parallel processing units, then performance is improved, but device complexity increases
Solution Approach 1:
The patent segments the deep learning library into multiple independent optimization passes, each handling specific optimization opportunities. This allows comprehensive optimizations to be applied in a modular fashion, improving performance while keeping each individual optimization component manageable and maintainable.
Solution Approach 2:
The patent performs optimization analysis and code generation in preliminary passes before actual execution. By pre-computing optimization opportunities and generating optimized code representations in advance, the system achieves high performance without adding complexity to the runtime execution path.
2Adaptability or versatility
If the API is updated to support added functionality, then adaptability is improved, but ease of operation deteriorates
Solution Approach 1:
The patent implements a universal API design where a single function call can trigger multiple optimization passes and handle various operation types. This multi-functional approach allows the API to support added functionality without requiring users to learn or call multiple different functions, maintaining ease of operation while increasing adaptability.
Solution Approach 2:
The optimization system automatically analyzes computation graphs, identifies optimization opportunities, and generates optimized code without requiring explicit user configuration. The API serves itself by automatically applying appropriate optimizations based on the detected computation patterns, eliminating the need for users to manually configure complex optimization parameters.
3Manufacturing precision
If multiple optimization passes are implemented, then manufacturing precision is improved, but device complexity increases
Solution Approach 1:
The patent divides the optimization process into distinct sequential passes, each focused on specific optimization techniques. This segmentation allows each pass to be independently implemented, tested, and maintained, reducing the complexity burden of having multiple optimization passes while achieving high optimization precision through their combined effect.
Solution Approach 2:
The patent introduces intermediate representation structures that carry computation graph information between optimization passes. These intermediaries allow each optimization pass to work with standardized data structures, reducing the complexity of coordinating multiple passes while enabling precise optimizations through iterative refinement.
Data Source
AI summary
Apparatuses, systems, and techniques to improve neural network computations. In at least one embodiment, a deep neural network library receives computation descriptors from one or more users and generates an optimized execution plan comprising one or more optimized operations to facilitate neural network computing.