Graph-Based Data Augmentation Policy Evolution for Neural Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data augmentation techniques are limited in expressive power, manually designed, and do not adapt to the domain or model, leading to inefficiencies in training complex neural networks, particularly in evolutionary computation, and require separate search phases that increase computational cost.
Innovation Solution
An evolutionary system that evolves data augmentation policies using a graph-based approach with NEAT (Topology and Weight Evolving Artificial Neural Networks) to adapt to both domain and network architecture, automatically determining optimal hyperparameters and interleaving operations into the model training process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If manual data augmentation pipelines are used with fixed operations in prespecified order, then the implementation is simple, but the expressive power is insufficient to deal with complex datasets
Solution Approach 1:
The patent implements dynamic data augmentation pipelines where operations, their parameters, and ordering are automatically adapted based on the dataset characteristics and model requirements. The system evolves augmentation policies through training, transitioning from static fixed pipelines to dynamic adaptive pipelines that respond to the specific task at hand.
Solution Approach 2:
The system performs self-service by automatically designing and optimizing data augmentation pipelines without manual intervention. The framework autonomously selects operations, determines their parameters, and establishes their ordering based on performance feedback, eliminating the need for manual pipeline design while achieving high expressive power.
2Ease of operation
If data augmentation pipelines are manually designed with fixed hyperparameters, then the design process is straightforward, but the pipelines do not adapt to the domain at hand
Solution Approach 1:
The patent incorporates feedback mechanisms where the performance of the model on validation data informs the optimization of data augmentation policies. The system uses this feedback to iteratively improve the augmentation pipeline, adapting it to the specific domain and task requirements while maintaining ease of operation through automated optimization.
Solution Approach 2:
The system automatically adjusts hyperparameters of data augmentation operations based on domain characteristics and model performance. Instead of using fixed manually-determined parameters, the framework dynamically changes parameters such as transformation probabilities, magnitudes, and operation sequences to optimize for the specific domain.
3Quantity of substance
If large datasets with irrelevant data are used, then the data coverage is comprehensive, but the training time increases significantly
Solution Approach 1:
The patent extracts and applies only the most relevant data augmentation operations for each specific task, rather than applying all possible operations to all data. This selective approach removes unnecessary computational overhead while maintaining comprehensive data coverage, thereby reducing training time without sacrificing data utilization effectiveness.
Solution Approach 2:
The system applies data augmentation operations selectively and partially based on task requirements, rather than exhaustively applying all possible operations. This partial action approach avoids the excessive computation that would result from comprehensive application of all augmentation techniques to entire large datasets.
4Measurement precision
If separate search phases are implemented for data augmentation optimization, then the augmentation policy can be optimized, but the computational cost increases
Solution Approach 1:
The patent merges the data augmentation optimization process with the model training process into a unified framework. Instead of implementing separate search phases that would double the computational cost, the system jointly optimizes both the model parameters and augmentation policies during a single training phase, achieving precise optimization at reduced computational cost.
Solution Approach 2:
The framework serves multiple functions simultaneously: it trains the model, optimizes augmentation policies, and adapts to domain characteristics all in one unified process. This multi-functionality eliminates the need for separate dedicated search phases while achieving comprehensive optimization.
Data Source
AI summary
A process for evolving a data augmentation policy for application to sample data from a dataset for us in training a neural network to perform a predetermined task is described. An initial population of candidate data augmentation policy models, each model including multiple nodes which are distinct data augmentation operations and multiple edges which have weight values representing a probability related to action by a second node on input data from a first node. The models in the population are evaluated by applying to the sample data and at least partially training the neural network using the augmented sample dataset. A fitness is determined based on the results of the training and models are selected either as final policy models or for reproduction and repeating of the evolution process until a final model is selected.