Dissertation and Thesis

Permanent URI for this communityhttp://164.52.219.250:4000/handle/10263/2146

Browse

Filters

20242

Settings

Subject: search.filters.subject.Class imbalance

Search Results

Now showing 1 - 2 of 2
  • No Thumbnail Available
    Item
    Enhancing Confidence Calibration in Long-Tailed Recognition
    (Indian Statistical Institute, Kolkata, 2024-06) Jana, Sasanka
    Deep neural networks often struggle with heavily class-imbalanced training datasets. Recently, two-stage methods have been developed to separate representation learning from classifier learning, aiming to enhance performance. However, the crucial issue of miscalibration remains. To tackle this, we introduce novel methods to improve both calibration and performance in such scenarios. Recognizing that predicted probability distributions of classes are closely tied to the number of class instances, we propose label-aware smoothing with balanced softmax. This strategy tackles the issue of differing levels of over-confidence among different categories, thereby improving the learning process of classifiers. Furthermore, to counteract potential bias in the dataset between the two stages caused by different sampling techniques, we incorporate shifted batch normalization into the decoupling framework. The methods we suggest have set fresh standards on numerous prevalent longtailed recognition datasets such as CIFAR10-LT and CIFAR100-LT.
  • No Thumbnail Available
    Item
    Learning with Long-Tailed Noisy Labels
    (Indian Statistical Institute, Kolkata, 2024-06) Dey, Sarbajit
    Deep neural networks (DNNs) have shown exceptionally good performance in a variety of activities by using correctly labelled and ’good’ training datasets. These remarkable results, however, are mostly observed with datasets that are carefully controlled and precisely structured. Conversely, data obtained from real-world applications frequently encounter substantial problems that are not commonly found in these ’good’ datasets. Two common biases frequently found in real-world data are: (i) long-tailed class distribution, where a small number of classes have a significant number of instances while the rest have only a few, and (ii) label noise, which refers to inaccuracies and errors in the assigned data labels. When learning models are specifically built to address only one of these biases, either by focusing on the long-tailed nature of the data or on the noise in the labels, their performance declines when they come across data that has both long-tailed distribution and noisy labels, which is a very common occurrence in real-world applications. This work investigates the complex issue of learning from datasets with long-tailed label noise. In real-world problems such as autonomous driving, medical diagnosis, and large-scale user-generated content platforms, the data obtained frequently shows these properties. Therefore, it is essential to create strong learning algorithms that can successfully address both problems at the same time. Our objective is to study and make meaningful contributions to the progress of deep learning methods that can effectively handle real-world data difficulties while being robust and dependable. We study the current methods for handling these learning problems, focuse on their shortcomings and try to improve the same. We propose Median of Means for centroid estimation on a clean subset of the dataset. We then use the SFA framework and Semi supervised learning for classification task on imbalanced noisy labels.