Neural Network Training With Adversarial Images for Fairness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Trained neural networks are susceptible to small perturbations that can significantly degrade their performance, and their effectiveness depends on the similarity between the training dataset and real-world data, leading to inconsistent performance across different environments.
Innovation Solution
Enhance neural network training using adversarial dataset generation to introduce imperceptible differences and distributionally robust optimization to create a training dataset that mirrors real-world conditions, ensuring the network's fairness and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If neural networks are trained on standard datasets, then they achieve good performance on similar data, but they become susceptible to small perturbations and perform inconsistently across different environments
Solution Approach 1:
The patent applies preliminary action by generating adversarial examples before deployment and incorporating them into the training dataset. This pre-training exposure to perturbed images prepares the neural network to handle small perturbations and distribution shifts in real-world scenarios, thereby improving reliability and robustness without requiring changes to the network architecture or deployment environment.
2Measurement precision
If the training dataset closely matches real-world data, then the network performs well in that specific environment, but it fails to generalize to diverse environments
Solution Approach 1:
The patent applies parameter changes by systematically transforming training images through various perturbations including color shifts, brightness adjustments, and geometric transformations. These parameter changes create a more diverse training dataset that exposes the neural network to a wider range of environmental conditions, enabling it to generalize across different real-world scenarios while maintaining prediction accuracy.
3Reliability
If adversarial examples are added to the training dataset, then robustness to perturbations improves, but computational complexity and training time increase
Solution Approach 1:
The patent applies self-service by implementing an automated adversarial example generation system that integrates directly into the training pipeline. The system automatically generates perturbed images, incorporates them into the training dataset, and retrains the neural network without requiring manual intervention or complex external tools. This self-contained approach reduces computational overhead and simplifies the training process while maintaining robustness improvements.
Data Source
AI summary
A system is disclosed that includes a computer that includes a processor and a memory, the memory including instructions executable by the processor to input an image acquired by a sensor to a neural network to output a prediction regarding an object included in the image. The neural network can be trained based on (a) a distributed robust optimization that minimizes an expectation applied to probability distributions of loss functions to select training images that yield a solution with a selected uncertainty level and (b) generating additional input images based on adversarial images.