Robust Deep Metric Learning In Open World

As an important branch of machine learning,metric learning is widely used in many practical applications such as information retrieval,object recognition,and pose estimation.In recent years,with the rapid development of deep learning,deep metric learning combined with deep neural networks has gradually become the mainstream method in metric learning.Although existing deep metric learning methods have achieved significant performance,they rely heavily on large-scale high-quality annotated training data.However,in many practical applications,such large-scale high-quality annotated training data are not available.To better utilize low-quality and limited training data in open-world applications,the robustness and generalization ability of deep metric learning is still facing a challenge.This dissertation focuses on deep metric learning applied in the open world.To tackle the problems existing in the open-world training data,such as small scale,no manual annotation,and noisy labels,this dissertation modifies the deep metric learning framework via theoretical innovations.The main innovative works in this dissertation are summarized as follows:We propose a robust Bayesian inference-based deep metric learning method.The existing deep metric learning methods can only achieve theoretical performance on large-scale high-quality labeled data because they are not robust against the noisy labels and have poor generalization ability on small-scale training data.To the best of our knowledge,it is the first work for hierarchical Bayesian deep metric learning.Specifically,a hierarchical Bayesian formulation with a class of self-adaptive spike-and-slab priors is constructed,which can automatically capture the structure information of the model parameters.Compared with the conventional deep metric learning methods,this model can be well trained using the small labeled data and is more robust to the label noise.Thus,this model is more practically useful in open-world applications,where the large-scale labeled data are generally not available due to the expensive cost.We proposed a hyperbolic space-based unsupervised deep metric learning is proposed.Learning feature embedding directly from images without any human supervision is a very challenging and essential task in the field of computer vision and machine learning.Following the paradigm in a supervised manner,most existing unsupervised metric learning approaches mainly focus on binary similarity in Euclidean space.However,these methods cannot achieve promising performance in many practical applications,where the manual information is lacking and data exhibits non-Euclidean latent anatomy.To address this limitation,an unsupervised hyperbolic metric learning method based on hyperbolic space is proposed.It considers the natural hierarchies of data by taking advantage of hyperbolic metric learning and hierarchical clustering,which can effectively excavate richer similarity information beyond binary in modeling.Extensive experimental results on benchmark datasets demonstrate that our method achieves state-of-the-art performance compared with current unsupervised deep metric learning approaches.We propose an adaptive Hierarchical metric learning method.Most existing deep metric learning methods with binary similarity are sensitive to noisy labels,which are widely present in open-world data.Since these noisy labels often cause severe performance degradation,enhancing the robustness and generalization ability of deep metric learning is crucial.Therefore,an adaptive hierarchical metric learning method is proposed.It considers two noiseinsensitive information,i.e.,class-wise divergence and sample-wise consistency.Specifically,class-wise divergence can effectively excavate richer similarity information beyond binary in modeling by taking advantage of Hyperbolic metric learning,while sample-wise consistency can further improve the generalization ability of the model using contrastive augmentation.More importantly,an adaptive strategy is designed to integrate this information in a unified view.Extensive experimental results on benchmark datasets demonstrate that the proposed method achieves state-of-the-art performance compared with current deep metric learning approaches.We propose a negative correlation mining-based weakly-supervised contrastive learning method for learning with noisy labels,which combines with the deep metric learning method.Current popular methods for training DNNs with noisy labels mainly focus on directly filtering out samples with low confidence or repeatedly mining valuable information from lowconfident samples.However,they cannot guarantee the robust generalization of models due to the ignorance of useful information hidden in noisy data.To address this issue,a new robust learning method is proposed to leverage the negative correlations from the noisy data.Specifically,in label space,we exploit the weakly-augmented data to filter samples and adopt classification loss on strong augmentations of the selected sample set,which can preserve the training diversity.While in metric space,weakly-supervised contrastive learning is applied to excavate these negative correlations hidden in noisy data.Extensive experiments have validated the superiority of our approach over existing state-of-the-art methods.