Automated Tiny Deep Learning Model Creation via Multi-Objective Reward Tuning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for creating Deep Learning models for resource-constrained embedded devices are inefficient due to the complexity of balancing multiple hardware constraints like accuracy, latency, and model size, with existing multi-objective optimization techniques being static and not dynamically tunable.
Innovation Solution
A method and system for automated creation of tiny Deep Learning models using a multi-objective reward function, combining Fast Evolutionary Algorithm and Deep Q-Learning Network to iteratively refine Neural Architecture Search, allowing dynamic adjustment of weightages for accuracy, latency, and size based on hardware constraints.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multi-objective optimization techniques are used to balance accuracy, latency, and model size, then model performance on resource-constrained devices is improved, but the complexity of the optimization process increases and the methods become static and not dynamically tunable
Solution Approach 1:
The patent applies dynamics by making the reward function tunable through adjustable weightages for different performance metrics. The system allows dynamic adjustment of the reward function parameters to adapt to different hardware platforms and performance requirements, transforming a static optimization process into a dynamic one that can be reconfigured without retraining the models
Solution Approach 2:
The patent implements parameter changes by introducing configurable weightages in the multi-objective reward function that control the relative importance of accuracy, latency, and model size. These parameters can be modified to prioritize different objectives based on specific deployment scenarios, enabling flexible adaptation without changing the underlying model architecture
2Reliability
If existing multi-objective optimization methods are applied, then multiple performance metrics are balanced, but the methods lack adaptability to changing hardware constraints and requirements
Solution Approach 1:
The system enables dynamic adaptation to changing hardware constraints by allowing the reward function weightages to be adjusted based on target platform characteristics. This dynamic configuration allows the same optimization framework to adapt to different hardware scenarios without requiring methodological changes
Solution Approach 2:
The patent achieves universality by creating a multi-objective optimization framework that can handle multiple performance metrics (accuracy, latency, model size) simultaneously while adapting to different hardware platforms. The standardized interface with configurable parameters allows the same system to serve multiple deployment scenarios
3Reliability
If Neural Architecture Search is performed with multiple hardware constraints, then optimized models for specific platforms are obtained, but the search space becomes huge and computation time increases
Solution Approach 1:
The patent applies preliminary action by performing a coarse-grained search using Fast Evolutionary Algorithm to identify promising regions in the neural architecture search space before conducting fine-grained search. This preliminary filtering step reduces the search space significantly, allowing the subsequent detailed optimization to focus only on relevant candidates
Solution Approach 2:
The patent segments the Neural Architecture Search process into two distinct phases: coarse-grained search using Fast Evolutionary Algorithm for initial exploration and filtering, and fine-grained search using Deep Q-Learning for detailed optimization. This segmentation divides the huge search space into manageable segments that can be processed sequentially
Data Source
AI summary
State of art techniques existing method refer to handling multiple objectives such as accuracy and latency. However, the reward functions are static and not tunable at user end. Further, for NN search with hardware constraints, approaches combine various techniques such as Reinforcement learning, Evolutionary Algorithm (EA) etc., however hardly any work attempts to disclose combining different NAS approaches in unison to reduce the search space of other. Embodiments of the present disclosure provide a method and system for automated creation of tiny Deep Learning (DL) models to be deployed on a platform having a set of hardware constraints. The method performs a coarse-grained search using a Fast EA NAS model and then utilizes a fine-grained search to identify customized and optimized tiny model. The coarse-grained search and the fine-grained search performed by agents based on a weighted multi-objective reward function, which are tunable at user end.