Deep Learning from Noisy Image Labels with Quality Embedding.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Label noise in image datasets significantly degrades deep learning model performance.
• Existing methods using latent labels struggle with mismatches between latent and noisy labels.

Purpose of the Study:

• To propose a probabilistic model that explicitly accounts for the trustworthiness of noisy labels using a quality variable.
• To minimize the influence of label noise by identifying mismatches between latent and noisy labels.

Main Methods:

• Introduced a probabilistic model with a 'quality variable' to represent noisy label trustworthiness.
• Developed the Contrastive-Additive Noise (CAN) network with contrastive and additive layers.
• Derived an SGD algorithm with reparameterization tricks for scalable optimization.

Main Results:

• The proposed quality variable effectively minimizes the influence of label noise.
• CAN network successfully reduces the impact of label mismatches.
• The method demonstrates scalability to large datasets through an efficient SGD algorithm.

Conclusions:

• The novel probabilistic approach and CAN network offer a robust solution for deep learning with noisy image datasets.
• CAN outperforms state-of-the-art methods in handling label noise, enhancing visual recognition accuracy.