Robust Classification with Convolutional Prototype Learning

By Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, Chenglin Liu. 2018

464 citations recorded by the source corpus.

Convolutional neural networks (CNNs) have been widely used for image classification. Despite its high accuracies, CNN has been shown to be easily fooled by some adversarial examples, indicating that CNN is not robust enough for pattern classification. In this paper, we argue that the lack of robustness for CNN is caused by the softmax layer, which is a totally discriminative model and based on the assumption of closed world (i.e., with a fixed number of categories). To improve the robustness, we propose a novel learning framework called convolutional prototype learning (CPL). The advantage of using prototypes is that it can well handle the open world recognition problem and therefore improve the robustness. Under the framework of CPL, we design multiple classification criteria to train the network. Moreover, a prototype loss (PL) is proposed as a regularization to improve the intra-class compactness of the feature representation, which can be viewed as a generative model based on the Gaussian assumption of different classes. Experiments on several datasets demonstrate that CPL can achieve comparable or even better results than traditional CNN, and from the robustness perspective, CPL shows great advantages for both the rejection and incremental category learning tasks.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

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

Citations in the corpus

Continual Few-shot Relation Learning via Embedding Space Regularization and Data Augmentation (2022, background): “The Lpm loss penalizes the cases where the similarity score of the closest wrong label is higher than the score of the correct label (Yang et al., 2018).”
A Two-phase Prototypical Network Model for Incremental Few-shot Relation Classification (2020, methodology): “, 2019a) and encoded by prototypical network (Yang et al., 2018).”
Semantic Drift Compensation for Class-Incremental Learning (2020, methodology): “Furthermore, deep embeddings were reported to be superior to classification networks for transfer learning [34] and preliminary results suggest that they might be less prone to catastrophic forgetting [45].”
Prototype-based Neural Network Layers: Incorporating Vector Quantization (2018, methodology): “In [57], a fused network was applied to train a network on MNIST and Cifar10.”
Rethinking Semantic Segmentation: A Prototype View (2022, background): “, ||i−pm||) [123], yet other proximities can be applied.”
Incremental Event Detection via Knowledge Consolidation Networks (2020, methodology): “Inspired by the recent works about prototype learning (Snell et al., 2017; Yang et al., 2018) which refer to the prototype as the class representative point in feature space, we devise a prototype-based selection algorithm.”
Hyperspherical Prototype Networks (2019, unknown): “Prototypebased networks result in a simple output layout [45] and generalize quickly to new classes [11, 37, 46].”
Learning to Classify With Incremental New Class (2021, unknown): “9) CPL [30]: Uses the same structure network as LC-INC for embedding, and adopts distance-based cross-entropy loss to improve the intraclass compactness of the feature.”
Adversarial Reciprocal Points Learning for Open Set Recognition (2021, methodology): “As shown in Table 3, the ARPL outperforms training by three methods for in-distribution data only; the methods are cross-entropy loss (Baseline), GCPL [3] and RPL [8].”
Meta-Neighborhoods (2019, methodology): “Cosine similarity classifiers have been adopted to replace dot product classifiers for few-shot learning [8, 2] and robust classification [40].”
Attribute Prototype Network for Zero-Shot Learning (2020, background): “Prototype learning is considered to be more robust when handling open-set recognition [46, 32], out-of-distribution samples [3] and few-shot learning [33, 11, 28].”