Striking the Right Balance With Uncertainty

By Salman Hameed Khan, Munawar Hayat, Waqas Zamir, Jianbing Shen, Ling Shao. 2019

198 citations recorded by the source corpus.

Learning unbiased models on imbalanced datasets is a significant challenge. Rare classes tend to get a concentrated representation in the classification space which hampers the generalization of learned boundaries to new test examples. In this paper, we demonstrate that the Bayesian uncertainty estimates directly correlate with the rarity of classes and the difficulty level of individual samples. Subsequently, we present a novel framework for uncertainty based class imbalance learning that follows two key insights: First, classification boundaries should be extended further away from a more uncertain (rare) class to avoid over-fitting and enhance its generalization. Second, each sample should be modeled as a multi-variate Gaussian distribution with a mean vector and a covariance matrix defined by the sample's uncertainty. The learned boundaries should respect not only the individual samples but also their distribution in the feature space. Our proposed approach efficiently utilizes sample and class uncertainty information to learn robust features and more generalizable classifiers. We systematically study the class imbalance problem and derive a novel loss formulation for max-margin learning based on Bayesian uncertainty measure. The proposed method shows significant performance improvements on six benchmark datasets for face verification, attribute prediction, digit/object classification and skin lesion detection.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Probabilistic & Bayesian Methods

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Probabilistic & Bayesian Methods

Citations in the corpus

Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation (2020, methodology): “, meta-learning [63, 29, 32], causal inference [58], Bayesian methods [34], etc.”
Decoupling Representation and Classifier for Long-Tailed Recognition (2019, background): “The loss can vary at class-level for matching a given data distribution and improving the generalization of tail classes (Cui et al., 2019; Khan et al., 2017; Cao et al., 2019; Khan et al., 2019; Huang et al., 2019).”
Class-Balanced Loss Based on Effective Number of Samples (2019, background): “It is noteworthy to mention other efforts in dealing with data imbalance, including transferring the knowledge learned frommajor classes to minor classes [3, 32, 44, 6, 45] and designing a better training objective via metric learning [17, 48, 46] or Bayesian uncertainty estimates [22].”
Rethinking the Value of Labels for Improving Class-Imbalanced Learning (2020, background): “This posts a major challenge for modern deep learning frameworks [2,5,10,20,53], where even with specialized techniques such as data re-sampling approaches [2,5,41] or class-balanced losses [7,13,26], significant performance drops still remain under extreme class imbalance.”
Data Uncertainty Learning in Face Recognition (2020, background): “In face recognition task, several works have been proposed to leverage model uncertainty for analysis and learning of face representations [13, 51, 23].”
iSAID: A Large-scale Dataset for Instance Segmentation in Aerial Images (2019, background): “Such a class imbalance usually exists in both natural and aerial imagery datasets and it is important for real-world applications [13].”
Uncertainty-Aware Learning for Zero-Shot Semantic Segmentation (2020, background): “[34] aim at balanced learning from training data with both model and data uncertainty based on Monte Carlo sampling, which however incurs high computational costs.”
Multi-label Contrastive Predictive Coding (2020, background): “[8] investigate re-weighting approaches that encourages large margins to the decision boundary for minority classes; such a context is also studied for detection [31] and segmentation [25] where class imbalance exists.”
A Review of Uncertainty Quantification in Deep Learning: Techniques, Applications and Challenges (2020, background): “[231] yielded that the difficulty level of individual”
Orthogonal Projection Loss (2021, unknown): “…calculation to limit the variance and range of logits during training [64], constraining the feature space to follow speciﬁc distributions [53], and heuristically altering margins targeting spe-ciﬁc tasks like few-shot learning [31, 28], class imbalance [20, 21, 17, 30] and zero-shot learning [39].”
Long-tail learning via logit adjustment (2020, background): “for τ > 0, where νy = P(y) in Kim and Kim [2019], Ye et al.”
Exploring Balanced Feature Spaces for Representation Learning (2021, methodology): “, 2018) or the loss for each class (Cui et al., 2019; Khan et al., 2017; Cao et al., 2019; Khan et al., 2019) by reweighting.”
Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss (2019, background): “Uneven margins for imbalanced datasets are also proposed and studied in [32] and the recent work [25, 33].”
Distribution Alignment: A Unified Framework for Long-tail Visual Recognition (2021, unknown): “Class-level methods typically re-weight the standard loss with category-specific coefficients correlated with the sample distributions [18, 9, 3, 21, 20, 34].”
Distribution-Balanced Loss for Multi-Label Classification in Long-Tailed Datasets (2020, background): “example difficulty [22], estimated Bayesian uncertainty [20], gradient direction [30].”
Balanced Knowledge Distillation for Long-tailed Learning (2021, background): “Cost-sensitive re-weighting [10, 47, 23] assigns different weights to the classification loss terms corresponding to different classes [7, 23, 5] or different samples [29, 22, 27].”
Probabilistic Face Embeddings (2019, background): “Specific to face recognition, some studies [6, 16, 45] have leveraged the model uncertainty for analysis and learning of face representations, but to our knowledge, ours is the first work that utilizes data uncertainty(3) for recognition tasks.”