Neural Architecture Search for DNN Model Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The manual design of Deep Neural Network (DNN) models for various tasks and hardware configurations is inefficient, leading to sub-optimal performance, redundant deployment cycles, and incompatibility issues due to the complexity of Artificial Intelligence problems and varying hardware specifications.
Innovation Solution
A Neural Architecture Search (NAS) method that identifies tasks, estimates performance parameters, determines hardware parameters, selects optimal neural blocks, and generates an optimized DNN model by modifying unsupported operations into supported ones or using universal approximators, allowing for a globally applicable abstract parameterized deep network that can be instantiated at deployment time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If manual design of DNN architecture is performed for each device and task, then architecture can be customized to hardware specifications, but engineering effort and time consumption increase significantly
Solution Approach 1:
The patent segments the DNN architecture into reusable neural blocks that can be independently selected and combined. Each neural block is a modular unit with specific functionality, allowing the system to construct different architectures by composing these standardized blocks rather than designing from scratch for each device.
Solution Approach 2:
The patent creates a universal library of neural blocks that can serve multiple hardware platforms and tasks. These blocks are designed with universal interfaces and functionalities that make them applicable across different device architectures (CPU, GPU, NPU) and task types, eliminating the need for device-specific custom designs.
2Reliability
If separate architecture learning pipelines are created for each device and task, then optimization can be tailored to specific hardware, but redundant deployment cycles and failures occur
Solution Approach 1:
The patent performs preliminary actions by pre-defining and pre-testing neural blocks in a standardized library before actual deployment. The blocks are designed with built-in compatibility checks and fallback mechanisms, so when deploying to a new device, the system only needs to select appropriate blocks rather than undergoing full retraining and testing cycles.
Solution Approach 2:
The patent implements feedback mechanisms during the neural block selection process where the system evaluates hardware specifications and task requirements, then selects blocks that are most likely to succeed. This feedback loop prevents deployment of incompatible architectures before execution, reducing failures.
3Productivity
If unsupported operations are used in DNN model, then model performance can be optimized for specific tasks, but incompatibility issues arise with certain hardware specifications
Solution Approach 1:
The patent applies parameter changes by detecting unsupported operations and automatically substituting them with alternative operations that are supported by the target hardware. The system changes the operational parameters of the neural blocks to match hardware capabilities while maintaining functional equivalence, ensuring compatibility without sacrificing performance.
Solution Approach 2:
The patent introduces intermediary layers or conversion operations that act as mediators between the idealized DNN operations and the hardware-specific implementations. These intermediaries translate unsupported operations into supported equivalents, enabling seamless operation across different hardware platforms.
4Power
If Leaky ReLU and other vendor-specific operations are used, then computational efficiency can be improved, but compatibility issues occur across different vendor NPUs
Solution Approach 1:
The patent designs neural blocks with universal operation sets that work across multiple vendor NPUs. Instead of optimizing for a single vendor's specific operations, the blocks use a standardized set of operations that can be executed on different hardware architectures, achieving multi-vendor compatibility while maintaining good performance through carefully selected universal operations.
Solution Approach 2:
The patent dynamically changes the operational parameters of neural blocks based on the target hardware's capabilities. When deploying to a specific vendor's NPU, the system adjusts the operations to match that vendor's supported set, replacing vendor-specific operations like Leaky ReLU with compatible alternatives while preserving computational efficiency through optimized parameter settings.
Data Source
AI summary
Embodiments herein provide a NAS method of generating an optimized DNN model for executing a task in an electronic device. The method includes identifying the task to be executed in the electronic device. The method includes estimating a performance parameter to be achieved while executing the task. The method includes determining hardware parameters of the electronic device required to execute the task based on the performance parameter and the task, and determining optimal neural blocks from a plurality of neural blocks based on the performance parameter and the hardware parameter of the electronic device. The method includes generating the optimized DNN model for executing the task based on the optimal neural blocks, and executing the task using the optimized DNN model.