Uncertainty handling in convolutional neural networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) performance degrades with noisy data, particularly during testing.
• Existing methods struggle to effectively mitigate the impact of data indeterminacy in image classification tasks.

Purpose of the Study:

• To propose a novel Neutrosophic Convolutional Neural Network (NCNN) structure for robust image classification.
• To address the challenge of noisy data and data indeterminacy in neural network performance.

Main Methods:

• Images are mapped from the pixel domain to the neutrosophic (NS) domain (True, Indeterminacy, False sets).
• A two-path NCNN architecture is constructed, processing True and Indeterminacy sets in parallel.
• Weights are updated using backpropagation, and the NS concept is applied to CNN and VGG-Net.

Main Results:

• NCNN outperforms standard CNN by 5.11% and 2.21% on noisy MNIST and CIFAR-10 datasets.
• NVGG-Net shows accuracy improvements of 3.09% and 2.57% over VGG-Net on CIFAR-10 and CIFAR-100 datasets.
• The proposed method demonstrates significant robustness against various levels of Gaussian noise.

Conclusions:

• The two-path NCNN effectively handles noisy data and data indeterminacy, outperforming traditional CNNs.
• The neutrosophic approach offers a promising direction for enhancing the resilience of deep learning models in image classification.
• NCNN and NVGG-Net provide improved accuracy and robustness, especially in the presence of significant data noise.