Mitigating Embedding and Class Assignment Mismatch in Unsupervised Image Classification

Han, Sungwon; Park, Sung Won; Kim, Sundong; Cha, Meeyoung

doi:10.1007/978-3-030-58586-0_45

Cited by 25 publications

(26 citation statements)

References 18 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…It is important to note that certain deep image clustering methods report superior performance to our method [21,15,24,16]. However, most approaches are multi-stage methods consisting of initialization methods, multiple losses, and fine-tuning methods.…”

Section: Clustering Image Datasetsmentioning

confidence: 85%

“…k-means [7] can also be thought of as a special case of GMM where the clusters are represented with untilted spheres. Since these traditional clustering algorithms are limited to hand-designed parameters hence limited representability, there have been efforts to incorporate neural networks to model advanced clusters [4,12,13,14,15,16].…”

Section: Related Workmentioning

confidence: 99%

“…The first one is learning a prior task that enables the network to extract important features [2,18,19,21,22]. The clustering is either conducted in the learned abstract space partially or fully using standard algorithms [2,18,19] or other clustering objectives as loss functions to train the network [16,21,22]. Learning a self-supervised pre-text task [18,19,21,22] often enables learning a rich semantic representation that acts as a solid prior for the clustering.…”

Section: Related Workmentioning

confidence: 99%

See 2 more Smart Citations

Neural Mixture Models with Expectation-Maximization for End-to-end Deep Clustering

Tissera¹,

Vithanage²,

Wijesinghe³

et al. 2021

Preprint

View full text Add to dashboard Cite

Any clustering algorithm must synchronously learn to model the clusters and allocate data to those clusters in the absence of labels. Mixture model-based methods model clusters with predefined statistical distributions and allocate data to those clusters based on the cluster likelihoods. They iteratively refine those distribution parameters and member assignments following the Expectation-Maximization (EM) algorithm. However, the cluster representability of such hand-designed distributions that employ a limited amount of parameters is not adequate for most real-world clustering tasks. In this paper, we realize mixture model-based clustering with a neural network where the final layer neurons, with the aid of an additional transformation, approximate cluster distribution outputs. The network parameters pose as the parameters of those distributions. The result is an elegant, much-generalized representation of clusters than a restricted mixture of hand-designed distributions. We train the network end-to-end via batch-wise EM iterations where the forward pass acts as the E-step and the backward pass acts as the M-step. In image clustering, the mixture-based EM objective can be used as the clustering objective along with existing representation learning methods. In particular, we show that when mixture-EM optimization is fused with consistency optimization, it improves the sole consistency optimization performance in clustering. Our trained networks outperform single-stage deep clustering methods that still depend on k-means, with unsupervised classification accuracy of 63.8% in STL10, 58% in CIFAR10, 25.9% in CIFAR100, and 98.9% in MNIST.

show abstract

Section: Clustering Image Datasetsmentioning

confidence: 85%

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Neural Mixture Models with Expectation-Maximization for End-to-end Deep Clustering

Tissera¹,

Vithanage²,

Wijesinghe³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…As a solution in this paper, we introduce Elsa (Energy based learning for semi-supervised anomaly detection), a novel anomaly detection method that unifies contrastive learning and energy-based functions. Our model benefits from the high representation power of unsupervised contrastive learning via its pre-training step, which can accommodate existing algorithms [6,14,15]. It then applies a carefully designed energy function over the pre-trained embedding to learn the probability distribution p(x) of normal data, with the help of a small set of labels that indicate whether given samples are normal or OOD.…”

Section: Introductionmentioning

confidence: 99%

Elsa: Energy-based learning for semi-supervised anomaly detection

Han¹,

Song²,

Lee³

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

Anomaly detection aims at identifying deviant instances from the normal data distribution. Many advances have been made in the field, including the innovative use of unsupervised contrastive learning. However, existing methods generally assume clean training data and are limited when the data contain unknown anomalies. This paper presents Elsa, a novel semi-supervised anomaly detection approach that unifies the concept of energy-based models with unsupervised contrastive learning. Elsa instills robustness against any data contamination by a carefully designed fine-tuning step based on the new energy function that forces the normal data to be divided into classes of prototypes. Experiments on multiple contamination scenarios show the proposed model achieves SOTA performance. Extensive analyses also verify the contribution of each component in the proposed model. Beyond the experiments, we also offer a theoretical interpretation of why contrastive learning alone cannot detect anomalies under data contamination.

show abstract

“…Entropy-based balancing has often been Figure 1: Illustration for this work's basic concept: robust learning approach via clean sample selection using pseudolabels from unsupervised clustering algorithm. adopted to prevent degenerate solutions along with these kinds of objectives [16,21,41].…”

Section: Introductionmentioning

confidence: 99%

Improving Unsupervised Image Clustering With Robust Learning

Park¹,

Han²,

Kim³

et al. 2020

Preprint

Self Cite

View full text Add to dashboard Cite

Unsupervised image clustering methods often introduce alternative objectives to indirectly train the model and are subject to faulty predictions and overconfident results. To overcome these challenges, the current research proposes an innovative model RUC that is inspired by robust learning. RUC's novelty is at utilizing pseudo-labels of existing image clustering models as a noisy dataset that may include misclassified samples. Its retraining process can revise misaligned knowledge and alleviate the overconfidence problem in predictions. This model's flexible structure makes it possible to be used as an add-on module to state-of-the-art clustering methods and helps them achieve better performance on multiple datasets. Extensive experiments show that the proposed model can adjust the model confidence with better calibration and gain additional robustness against adversarial noise.

show abstract

Mitigating Embedding and Class Assignment Mismatch in Unsupervised Image Classification

Cited by 25 publications

References 18 publications

Neural Mixture Models with Expectation-Maximization for End-to-end Deep Clustering

Neural Mixture Models with Expectation-Maximization for End-to-end Deep Clustering

Elsa: Energy-based learning for semi-supervised anomaly detection

Improving Unsupervised Image Clustering With Robust Learning

Contact Info

Product

Resources

About