PROJECT TITLE :

Deep Learning From Noisy Image Labels With Quality Embedding

ABSTRACT:

In many visual identification tasks, the use of noisy picture datasets is on the rise. Deep learning algorithms are substantially hampered by label noise in datasets. The latent label, a recent trend, has demonstrated promising results in network designs by reducing the amount of label noise. The mismatch between latent labels and noisy labelling still has an impact on the predictions in these systems. For this problem, we suggest a probabilistic model in which an additional variable, called the quality variable, is included to indicate the reliability of noisy labels. With the use of subspace embedding, we are able to find the mismatch between latent and noisy labels, effectively reducing the impact of label noise. However, training can still make use of trusted labels. By using a contrastive-additive noise network (CAN), which has two key layers: a quality variable estimation in embedding space and an additive layer that aggregates prior predictions and label noise as posteriors, we can quickly get a working model that can be applied to real-world data with minimal effort. Our technique is also scalable to large datasets because to the reparameterization tricks we use in our SGD algorithm. Using a variety of noisy image datasets, we test the proposed technique. Results show that CAN outperforms the best deep learning techniques currently available.