In this paper, we introduce Random Erasing, a new data augmentation method for training the convolutional neural network (CNN). training, Erasing randomly selects rectangle region in an image and erases its pixels with random values. process, images various levels of occlusion are generated, which reduces risk over-fitting makes model robust to occlusion. is parameter learning free, easy implement, can be integrated most CNN-based recognition models. Albeit simple, complementary commonly...

Person re-identification (re-ID) models trained on one domain often fail to generalize well another. In our attempt, we present a "learning via translation" framework. the baseline, translate labeled images from source target in an unsupervised manner. We then train re-ID with translated by supervised methods. Yet, being essential part of this framework, image-image translation suffers information loss source-domain labels during translation. Our motivation is two-fold. First, for each...

This paper proposed a Soft Filter Pruning (SFP) method to accelerate the inference procedure of deep Convolutional Neural Networks (CNNs). Specifically, SFP enables pruned filters be updated when training model after pruning. has two advantages over previous works: (1) Larger capacity. Updating previously provides our approach with larger optimization space than fixing zero. Therefore, network trained by capacity learn from data. (2) Less dependence on pretrained model. Large train scratch...

Unsupervised Domain Adaptation (UDA) makes predictions for the target domain data while manual annotations are only available in source domain. Previous methods minimize discrepancy neglecting class information, which may lead to misalignment and poor generalization performance. To address this issue, paper proposes Contrastive Network (CAN) optimizing a new metric explicitly models intra-class inter-class discrepancy. We design an alternating update strategy training CAN end-to-end manner....

In this paper, we introduce Random Erasing, a new data augmentation method for training the convolutional neural network (CNN). training, Erasing randomly selects rectangle region in an image and erases its pixels with random values. process, images various levels of occlusion are generated, which reduces risk over-fitting makes model robust to occlusion. is parameter learning free, easy implement, can be integrated most CNN-based recognition models. Albeit simple, complementary commonly...

Industrial recommender systems have embraced deep learning algorithms for building intelligent to make accurate recommendations. At its core, offers powerful ability representations from data, especially user and item representations. Existing learning-based models usually represent a by one representation vector, which is insufficient capture diverse interests large-scale users in practice. In this paper, we approach the of different view, representing with multiple vectors encoding aspects...

Semi-supervised image classification aims to classify a large quantity of unlabeled images by typically harnessing scarce labeled images. Existing semi-supervised methods often suffer from inadequate accuracy when encountering difficult yet critical images, such as outliers, because they treat all equally and conduct classifications in an imperfectly ordered sequence. In this paper, we employ the curriculum learning methodology investigating difficulty classifying every image. The...

Deeper and wider convolutional neural networks (CNNs) achieve superior performance but bring expensive computation cost. Accelerating such overparameterized network has received increased attention. A typical pruning algorithm is a three-stage pipeline, i.e., training, pruning, retraining. Prevailing approaches fix the pruned filters to zero during retraining and, thus, significantly reduce optimization space. Besides, they directly prune large number of at first, which would cause...

In this paper, we investigate Universal Domain Adaptation (UniDA) problem, which aims to transfer the knowledge from source target under unaligned label space. The main challenge of UniDA lies in how separate common classes (i.e., shared across domains), private only exist one domain). Previous works treat samples as generic class but ignore their intrinsic structure. Consequently, resulting representations are not compact enough latent space and can be easily confused with samples. To...

Unsupervised domain adaptation (UDA) makes predictions for the target data while manual annotations are only available in source domain. Previous methods minimize discrepancy neglecting class information, which may lead to misalignment and poor generalization performance. To tackle this issue, paper proposes contrastive network (CAN) that optimizes a new metric named Contrastive Domain Discrepancy explicitly modeling intra-class inter-class discrepancy. optimize CAN, two technical issues...

Few-shot segmentation aims to train a model that can fast adapt novel classes with few exemplars. The conventional training paradigm is learn make predictions on query images conditioned the features from support images. Previous methods only utilized semantic-level prototypes of as conditional information. These cannot utilize all pixel-wise information for predictions, which however critical task. In this paper, we focus utilizing relationships between and facilitate few-shot We design...

The goal of continual learning is to improve the performance recognition models in sequentially arrived data. Although most existing works are established on premise from scratch, growing efforts have been devoted incorporating benefits pre-training. However, how adaptively exploit pre-trained knowledge for each incremental task while maintaining its generalizability remains an open question. In this work, we present extensive analysis a model (CLPM), and attribute key challenge progressive...

This paper proposed a Soft Filter Pruning (SFP) method to accelerate the inference procedure of deep Convolutional Neural Networks (CNNs). Specifically, SFP enables pruned filters be updated when training model after pruning. has two advantages over previous works: (1) Larger capacity. Updating previously provides our approach with larger optimization space than fixing zero. Therefore, network trained by capacity learn from data. (2) Less dependence on pre-trained model. Large train scratch...

Domain adaptive semantic segmentation aims to train a model performing satisfactory pixel-level predictions on the target with only out-of-domain (source) annotations. The conventional solution this task is minimize discrepancy between source and enable effective knowledge transfer. Previous domain minimization methods are mainly based adversarial training. They tend consider globally, which ignore pixel-wise relationships less discriminative. In paper, we propose build cycle association...

Open set domain adaptation aims to diminish the shift across domains, with partially shared classes. There exist unknown target samples out of knowledge source domain. Compared close setting, how separate (unshared) class from known (shared) ones plays key role. Whereas, previous methods did not emphasize semantic structure open data, which may introduce bias into alignment and confuse classifier around decision boundary. In this paper, we exploit data two aspects: 1) Semantic Categorical...

Continual learning aims to enable a model incrementally learn knowledge from sequentially arrived data. Previous works adopt the conventional classification architecture, which consists of feature extractor and classifier. The is shared across tasks or classes, but one specific group weights classifier corresponding new class should be expanded. Consequently, parameters continual learner gradually increase. Moreover, as contains all historical certain size memory usually required store...

Recent years have witnessed the success of deep neural networks in dealing with a plenty practical problems. Dropout has played an essential role many successful networks, by inducing regularization model training. In this paper, we present new regularized training approach: Shakeout. Instead randomly discarding units as does at stage, Shakeout chooses to enhance or reverse each unit's contribution next layer. This minor modification statistical trait: regularizer induced adaptively combines...

This paper focuses on regularizing the training of convolutional neural network (CNN). We propose a new regularization approach named ``PatchShuffle`` that can be adopted in any classification-oriented CNN models. It is easy to implement: each mini-batch, images or feature maps are randomly chosen undergo transformation such pixels within local patch shuffled. Through generating and with interior orderless patches, PatchShuffle creates rich variations, reduces risk overfitting, viewed as...

Recent years have witnessed the success of deep neural networks in dealing with a plenty practical problems. The invention effective training techniques largely contributes to this success. so-called "Dropout" scheme is one most powerful tool reduce over-fitting. From statistic point view, Dropout works by implicitly imposing an L2 regularizer on weights. In paper, we present new scheme: Shakeout. Instead randomly discarding units as does at stage, our method chooses enhance or inverse...

Unsupervised Domain Adaptation (UDA) makes predictions for the target domain data while manual annotations are only available in source domain. Previous methods minimize discrepancy neglecting class information, which may lead to misalignment and poor generalization performance. To address this issue, paper proposes Contrastive Network (CAN) optimizing a new metric explicitly models intra-class inter-class discrepancy. We design an alternating update strategy training CAN end-to-end manner....

The transformer architectures with attention mechanisms have obtained success in Nature Language Processing (NLP), and Vision Transformers (ViTs) recently extended the application domains to various vision tasks. While achieving high performance, ViTs suffer from large model size computation complexity that hinders deployment of them on edge devices. To achieve throughput hardware preserve accuracy simultaneously, we propose VAQF, a framework builds inference accelerators FPGA platforms for...

Industrial recommender systems usually consist of the matching stage and ranking stage, in order to handle billion-scale users items. The retrieves candidate items relevant user interests, while sorts by interests. Thus, most critical ability is model represent interests for either stage. Most existing deep learning-based models one as a single vector which insufficient capture varying nature user's In this paper, we approach problem from different view, with multiple vectors encoding...