Random Neural Architecture Search via Reinforcement Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manually designing deep neural networks is inefficient due to the vast search space of hyperparameters, and existing automated methods like neural architecture search and random graph neural networks lack flexibility in exploring effective architectures.
Innovation Solution
A method using reinforcement learning with a deep deterministic policy gradient (DDPG) to balance intra- and extra-connections in random graph neural networks, allowing for the discovery of efficient neural architectures by optimizing FLOPs and error rate through a combination of Watts-Strogatz graph construction and additional parameters for inter-block connections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manual design of neural networks is used, then domain expertise and architectural engineering skills can be applied, but the process is tedious and inefficient due to the vast search space of hyperparameters
Solution Approach 1:
The system enables automatic neural network architecture search where the algorithm autonomously explores the search space and discovers optimal architectures without requiring manual intervention. The reinforcement learning agent independently evaluates configurations and iteratively improves the network design, eliminating the need for tedious manual hyperparameter tuning while maintaining high design quality
Solution Approach 2:
The invention transforms the discrete, combinatorial architecture search problem into a continuous optimization problem by representing network architectures as continuous parameters. This allows the use of reinforcement learning with continuous action spaces to efficiently navigate the vast hyperparameter search space, dramatically improving design speed compared to traditional manual or discrete search methods
2Productivity
If neural architecture search is used to automate network design, then optimal architectures can be discovered, but the search space remains immense and intractable
Solution Approach 1:
The architecture search space is segmented into hierarchical levels, where the reinforcement learning agent makes decisions at different granularities (e.g., selecting network depth, then width, then specific layer configurations). This segmentation breaks down the intractable monolithic search space into manageable sub-spaces that can be explored efficiently through sequential decision-making
Solution Approach 2:
The invention introduces a continuous dimension to the traditionally discrete architecture search space. By representing architectural parameters as continuous variables rather than discrete choices, the search transforms from a combinatorial explosion problem into a continuous optimization landscape that reinforcement learning can navigate more efficiently using gradient-based methods
3Adaptability or versatility
If random graph neural networks are used, then manual architecture design constraints are relaxed, but flexibility in exploring effective architectures is reduced due to use of only internal connections
Solution Approach 1:
The system dynamically adjusts the connection structure of random graph neural networks during the reinforcement learning process. Instead of using fixed random graphs, the algorithm learns to modify and adapt the graph structures based on performance feedback, enabling both flexibility in exploration and effectiveness in discovering optimal architectures through iterative refinement
Solution Approach 2:
The reinforcement learning agent acts as an intermediary between the random graph structure and the final effective architecture. It takes the randomly initialized graphs as starting points and systematically modifies them through learned transformations, bridging the gap between random exploration and targeted optimization to achieve both flexibility and effectiveness
Data Source
AI summary
Disclosed herein is a system and method for novel neural architecture search using a random graph network backbone to facilitate the creation of an efficient network structure. The method utilizes reinforcement learning algorithms to build a complex relationship between intra-connections (i.e., links between blocks in a random graph network) and extra-connections (i.e., links among blocks across the random graphs network) for discovering an efficient random neural architecture.