Deep Learning Model Training via Hardware Emulation Fidelity Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in training deep learning models for autonomous vehicles lies in the impracticality of testing and validating on real-world roads due to cost and resource constraints, particularly with specialized DSP hardware, which is not readily available in large numbers, leading to the need for efficient emulation methods that balance fidelity and cost.
Innovation Solution
A method involving running a deep learning model on both specialized and generalized hardware platforms, with imperfect emulation, computing adjustments based on result differences, and applying corrections to ensure accuracy and efficiency, allowing for training on a cloud platform with varying levels of fidelity to match current compute resource costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If real-world roads are used for testing and validating deep learning models, then model accuracy and safety can be verified, but cost and resource constraints make this impractical
Solution Approach 1:
The patent creates a virtual copy of the specialized DSP hardware platform using emulation technology. This virtual replica allows extensive testing and validation of deep learning models without requiring physical specialized hardware or real-world road conditions, thereby reducing costs and resources while maintaining validation accuracy through computational simulation
Solution Approach 2:
The patent replaces physical mechanical testing systems (real-world roads, physical sensors, specialized DSP hardware) with a computational simulation system. By substituting the mechanical/physical testing environment with a software-based emulation and cloud computing platform, the system achieves scalable, cost-effective validation while maintaining model accuracy through virtual reproduction of hardware behavior
2Reliability
If specialized DSP hardware is used for training deep learning models, then model performance and safety can be ensured, but hardware availability and cost increase
Solution Approach 1:
The patent creates a virtual copy of the specialized DSP hardware platform using emulation technology. This virtual replica allows extensive testing and validation of deep learning models without requiring physical specialized hardware or real-world road conditions, thereby reducing costs and resources while maintaining validation accuracy through computational simulation
Solution Approach 2:
The patent employs a cloud-based virtualization platform that can simulate multiple hardware configurations and provide various levels of emulation fidelity. This universal platform can accommodate different training requirements and hardware emulations without requiring dedicated specialized hardware for each scenario, improving accessibility and reducing costs
3Productivity
If cloud platform with varying levels of fidelity is used for training, then cost efficiency is improved, but emulation accuracy may be compromised
Solution Approach 1:
The patent implements a dynamic fidelity adjustment mechanism that adapts the level of emulation accuracy based on the specific training requirements and computational resources available. The system can switch between different fidelity levels during the training process, using higher accuracy when needed and lower accuracy for cost-efficient bulk processing, thereby optimizing the balance between cost efficiency and accuracy
Solution Approach 2:
The patent changes the fidelity parameter of the emulation system dynamically during training. By adjusting the fidelity level as a configurable parameter, the system can optimize the trade-off between computational cost and emulation accuracy for different stages of model training, achieving cost efficiency without compromising necessary accuracy
Data Source
AI summary
A system may provide a method of training a deep learning (DL) model, comprising: running a first version of the DL model on a first hardware platform using a first input set; running a second version of the DL model on a second hardware platform using the first input set, comprising imperfectly emulating the first hardware platform on the second hardware platform; computing an adjustment based at least in part on a difference in results between the first version of the DL model and second version of the DL model; and training the DL model on the second hardware platform using the adjustment and a plurality of input sets.