Neural Network Image Classification With Correlation-Based Error Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks are susceptible to adversarial attacks and edge/corner cases, leading to inaccurate predictions, and retraining to address these issues is computationally expensive.
Innovation Solution
Implementing a combination of a neural network with a correlation structure, such as a Gaussian process, using a reference dataset to improve prediction accuracy by comparing input images with reference images to correct misclassifications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a neural network is used for image classification, then prediction capability is provided, but prediction accuracy deteriorates for edge cases, corner cases, and adversarial attacks
Solution Approach 1:
A correlation structure (Gaussian process) is introduced as an intermediary between the neural network and the final prediction. This correlation structure compares the input image with reference images to determine similarity, and uses this similarity information to adjust or correct the neural network's prediction, thereby improving accuracy for edge cases and adversarial attacks without retraining the network
Solution Approach 2:
Reference images are pre-collected and stored in a reference dataset before deployment. These reference images represent various edge cases, corner cases, and adversarial examples. During operation, the correlation structure performs preliminary comparison between the input image and these pre-prepared reference images to guide the final prediction
2Reliability
If the neural network is retrained to improve accuracy for edge cases and adversarial attacks, then prediction accuracy improves, but computational cost and training time increase
Solution Approach 1:
Instead of retraining the entire neural network, the system creates a copy or supplement in the form of a correlation structure that operates alongside the existing network. This correlation structure uses pre-computed reference images and similarity comparisons to correct predictions, avoiding the high computational cost of full network retraining while still improving accuracy
3Productivity
If the neural network operates with high confidence, then prediction speed is maintained, but false predictions increase for adversarial and edge cases
Solution Approach 1:
The correlation structure provides feedback to the neural network's prediction by comparing the input image with reference images. When the correlation structure detects high similarity between the input and a reference image (especially for edge cases or adversarial examples), it adjusts the prediction confidence or corrects the prediction, creating a feedback loop that improves reliability without significantly impacting speed
Data Source
AI summary
Systems and methods for performing image classification are disclosed. The methods include, by a processor: receiving an input image, generating a label prediction corresponding to the input image using a trained neural network, generating a correlation structure based on a comparison of the input image with each of a plurality of reference images, and generating an updated label prediction corresponding to the input image using the label prediction and the correlation structure.