Refining activation downsampling with SoftPool

Stergiou, Alexandros; Poppe, Ronald; Kalliatakis, Grigorios

doi:10.1109/iccv48922.2021.01019

Cited by 187 publications

(75 citation statements)

References 34 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…As a result, the model may fail to represent every attribute of an entity accurately. In practice, we adopt an exponentially weighted pooling method similar to softpool (Stergiou et al, 2021):…”

Section: Proposed Methodsmentioning

confidence: 99%

“…We also treat the known types of the central entity as its neighbors to use them to infer the missing types. To aggregate the inference results generated by N2T and Agg2T mechanism, we adopt a carefully designed pooling method similar to softpool (Stergiou et al, 2021). Experiments show that this pooling method can produce stable and interpretable inference results.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Context-aware Entity Typing in Knowledge Graphs

Pan¹,

Wei

Mao

2021

Preprint

View full text Add to dashboard Cite

Knowledge graph entity typing aims to infer entities' missing types in knowledge graphs which is an important but under-explored issue. This paper proposes a novel method for this task by utilizing entities' contextual information. Specifically, we design two inference mechanisms: i) N2T: independently use each neighbor of an entity to infer its type; ii) Agg2T: aggregate the neighbors of an entity to infer its type. Those mechanisms will produce multiple inference results, and an exponentially weighted pooling method is used to generate the final inference result. Furthermore, we propose a novel loss function to alleviate the false-negative problem during training. Experiments on two real-world KGs demonstrate the effectiveness of our method. The source code and data of this paper can be obtained from https://github.com/ CCIIPLab/CET.

show abstract

Section: Proposed Methodsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Context-aware Entity Typing in Knowledge Graphs

Pan¹,

Wei

Mao

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…For some datasets, the data imbalance problem can lead to a certain bias in the classification results of the deep learning model, i.e., it is biased towards the category with more data, and there are some methods for dealing with the data imbalance problem, such as oversampling and undersampling, etc. The method we use is SMOTE sampling [40], which is a method for constructing artificial samples, and SMOTE is an oversampling algorithm, which constructs new samples of small classes instead of generating copies of samples that already exist in the small classes, i.e., the algorithm constructs data that are new samples and do not exist in the original dataset. The basic idea is to find the K nearest neighbor samples according to the Euclidean distance for each small sample class, randomly select a sample point, and randomly select a point on the line segment between the sample point and the nearest neighbor sample point in the feature space.…”

Section: Swin Transformer Blockmentioning

confidence: 99%

Expression recognition method combining convolutional features and Transformer

Zhu¹,

Li²,

Sun³

2023

MFC

View full text Add to dashboard Cite

<p style='text-indent:20px;'>Expression recognition has been an important research direction in the field of psychology, which can be used in traffic, medical, security, and criminal investigation by expressing human feelings through the muscles in the corners of the mouth, eyes, and face. Most of the existing research work uses convolutional neural networks (CNN) to recognize face images and thus classify expressions, which does achieve good results, but CNN do not have enough ability to extract global features. The Transformer has advantages for global feature extraction, but the Transformer is more computationally intensive and requires a large amount of training data. So, in this paper, we use the hierarchical Transformer, namely Swin Transformer, for the expression recognition task, and its computational power will be greatly reduced. At the same time, it is fused with a CNN model to propose a network architecture that combines the Transformer and CNN, and to the best of our knowledge, we are the first to combine the Swin Transformer with CNN and use it in an expression recognition task. We then evaluate the proposed method on some publicly available expression datasets and can obtain competitive results.</p>

show abstract

“…Local Importance Pooling (LIP) [ 42 ] utilizes learned weights as a subnetwork attention-based mechanism. SoftPool [ 43 ] downsamples the activation map based on the exponentially weighted sum of the original pixels within the kernel region (softmax normalization). This can improve the representation of high-contrast regions and present around object edges or specific feature activations.…”

Section: Related Workmentioning

confidence: 99%

Discriminative Shape Feature Pooling in Deep Neural Networks

Dixit

2022

J. Imaging

View full text Add to dashboard Cite

Although deep learning approaches are able to generate generic image features from massive labeled data, discriminative handcrafted features still have advantages in providing explicit domain knowledge and reflecting intuitive visual understanding. Much of the existing research focuses on integrating both handcrafted features and deep networks to leverage the benefits. However, the issues of parameter quality have not been effectively solved in existing applications of handcrafted features in deep networks. In this research, we propose a method that enriches deep network features by utilizing the injected discriminative shape features (generic edge tokens and curve partitioning points) to adjust the network’s internal parameter update process. Thus, the modified neural networks are trained under the guidance of specific domain knowledge, and they are able to generate image representations that incorporate the benefits from both handcrafted and deep learned features. The comparative experiments were performed on several benchmark datasets. The experimental results confirmed our method works well on both large and small training datasets. Additionally, compared with existing models using either handcrafted features or deep network representations, our method not only improves the corresponding performance, but also reduces the computational costs.

show abstract

Refining activation downsampling with SoftPool

Cited by 187 publications

References 34 publications

Context-aware Entity Typing in Knowledge Graphs

Context-aware Entity Typing in Knowledge Graphs

Expression recognition method combining convolutional features and Transformer

Discriminative Shape Feature Pooling in Deep Neural Networks

Contact Info

Product

Resources

About