Neural Network Self-Supervised Labeling via Image Perturbation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The process of labeling training images for neural networks is time-consuming, costly, and often restricted by legal or contractual limitations, making it difficult to utilize crowd-sourcing for all types of images, especially in medical or proprietary contexts.
Innovation Solution
The method involves determining object class probabilities for pixels in input images using neural networks, applying image perturbations to improve object recognition accuracy without additional labeled data, and using generative adversarial networks to generate distribution-based perturbations that enhance the neural network's performance in automated object recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If crowd-sourcing approaches are used to label training images, then the time and cost involved in labeling is reduced, but this approach cannot be applied to images subject to legal, contractual, or privacy restrictions
Solution Approach 1:
The system enables self-service by having the neural network automatically generate labels for training images through self-supervised learning. The model predicts labels for images and uses these predictions to train itself, eliminating the need for manual crowd-sourcing while maintaining versatility across different image types including those with legal or privacy restrictions.
2Reliability
If extensive labeled training data is collected and labeled manually, then the neural network can be trained effectively, but the process becomes time-consuming, costly, and restricted by legal or contractual limitations
Solution Approach 1:
The neural network performs self-service by automatically generating its own training labels through predictive modeling. The system processes images, generates probability distributions for different labels, and uses these self-generated labels for training, thereby achieving effective training without manual labeling time and costs.
Solution Approach 2:
The system performs preliminary action by pre-computing label predictions and probability distributions for training images before actual training occurs. These pre-computed predictions serve as the foundation for self-supervised training, enabling the network to learn from unlabeled data without waiting for manual annotation.
3Measurement precision
If manual labeling of training images is performed to ensure accuracy, then the training data quality is high, but the process is laborious and cannot be applied to all image types
Solution Approach 1:
The system achieves both precision and productivity through self-service automation. The neural network generates high-quality labels automatically by computing probability distributions across multiple labels and selecting the most probable labels, maintaining accuracy while dramatically increasing productivity compared to manual labeling processes.
Solution Approach 2:
The patent replaces the mechanical process of manual human labeling with an automated computational system. The neural network uses algorithmic processes to generate labels based on image analysis and probability calculations, substituting human labor with automated mechanics that can process images rapidly and consistently without legal or contractual restrictions.
Data Source
AI summary
A method includes determining object class probabilities of pixels in a first input image by examining the first input image in a forward propagation direction through layers of artificial neurons of an artificial neural network. The object class probabilities indicate likelihoods that the pixels represent different types of objects in the first input image. The method also includes selecting, for each of two or more of the pixels, an object class represented by the pixel by comparing the object class probabilities of the pixels with each other, determining an error associated with the object class that is selected for each pixel of the two or more pixels, determining one or more image perturbations by back-propagating the errors associated with the object classes selected for the pixels of the first input image through the layers of the neural network without modifying the neural network, and modifying a second input image by applying the one or more image perturbations to one or more of the first input image or the second input image prior to providing the second input image to the neural network for examination by the neurons in the neural network for automated object recognition in the second input image.