In this paper, we propose a novel training procedure for the continual
representation learning problem in which a neural network model is sequentially
learned to alleviate catastrophic forgetting in visual search tasks. Our
method, called Contrastive Supervised Distillation (CSD), reduces feature
forgetting while learning discriminative features. This is achieved by
leveraging labels information in a distillation setting in which the student
model is contrastively learned from the teacher model. Extensive experiments
show that CSD performs favorably in mitigating catastrophic forgetting by
outperforming current state-of-the-art methods. Our results also provide
further evidence that feature forgetting evaluated in visual retrieval tasks is
not as catastrophic as in classification tasks. Code at:
https://github.com/N iccoBiondi/Contrasti veSupervisedDistilla tion.
Media Integration and Communication Center (MICC), Dipartimento di Ingegneria
メディア統合通信センター(MICC, Dipartimento di Ingegneria)
0.74
dell’Informazione, Universit`a degli Studi di Firenze
dell'Informazione, Universit`a degli Studi di Firenze
0.42
name.surname@unifi.i t
name.surname@unifi.i t
0.29
Abstract. In this paper, we propose a novel training procedure for the continual representation learning problem in which a neural network model is sequentially learned to alleviate catastrophic forgetting in visual search tasks.
This is achieved by leveraging labels information in a distillation setting in which the student model is contrastively learned from the teacher model.
Our results also provide further evidence that feature forgetting evaluated in visual retrieval tasks is not as catastrophic as in classification tasks.
1 Introduction Deep Convolutional Neural Networks (DCNNs) have significantly advanced the field of visual search or visual retrieval by learning powerful feature representations from data [1,2,3].
Current methods predominantly focus on learning feature representations from static datasets in which all the images are available during training [4,5,6].
This operative condition is restrictive in real-world applications since new data are constantly emerging and repeatedly training DCNN models on both old and new images is timeconsuming.
Static datasets, typically stored on private servers, are also increasingly problematic because of the societal impact associated with privacy and ethical issues of modern AI systems [7,8].
These problems may be significantly reduced in incremental learning scenarios as the computation is distributed over time and training data are not required to be stored on servers.
The challenge of learning feature representation in incremental scenarios has to do with the inherent problem of catastrophic forgetting, namely the loss of previously learned knowledge when new knowledge is assimilated [9,10].
catastrophic forgetting is typically observed by a clear reduction in classification accuracy [11,12,13,14,15].
破滅的な忘れ物は通常、明確な分類精度の低下[11,12,13,14,15]によって観察される。
0.44
The fundamental differences with respect to learning internal feature representation for visual search tasks are: (1) evaluation metrics do not use classification accuracy (2) visual search data have typically a finer granularity with respect to categorical data and (3) no classes are required to be specifically learned.
These differences might suggest different origins of the two catastrophic forgetting phenomena.
これらの違いは、2つの破滅的な忘れる現象の異なる起源を示唆するかもしれない。
0.50
In this regard, some recent works provide some evidence showing the importance of the specific task when evaluating the catastrophic forgetting of the learned representations [16,17,18,19].
We argue that such evidence is relevant in visual search tasks and that it can be exploited with techniques that learn incrementally without storing past samples in a memory buffer [20].
According to this, in this paper, we propose a new distillation method for the continual representation learning task, in which the search performance degradation caused by feature forgetting is jointly mitigated while learning discriminative features.
We follow the basic working principle of contrastive loss [21] used in self-supervised learning, to effectively leverage label information in a distillation-based training procedure in which we replace anchor features with the feature of the teacher model.
1. We address the problem of continual representation learning proposing a novel method that leverages label information in a contrastive distillation learning setup.
2. Experimental results on different benchmark datasets show that our CSD training
2. 異なるベンチマークデータセットによる実験結果から、我々のCSDトレーニングが示される
0.68
procedure achieves state-of-the-art performance.
プロシージャは最先端のパフォーマンスを達成する。
0.31
3. Our results confirm that feature forgetting in visual retrieval using fine-grained
3. きめ細かな粒度を用いた視覚検索における特徴忘れの確認
0.60
datasets is not as catastrophic as in classification.
データセットは分類ほど壊滅的ではない。
0.76
2 Related Works Continual Learning (CL).
関連作品2件 継続学習(CL)。
0.75
CL has been largely developed in the classification setting, where methods have been broadly categorized based on exemplar [22,23,24,25] and regularization [26,27,20,28].
Only recently, continual learning for feature representation is receiving increasing attention and few works pertinent to the regularizationbased category has been proposed [17,18,19].
The work in [17] proposed an unsupervised alignment loss between old and new feature distributions according to the Mean Maximum Discrepancy (MMD) distance [29].
The work [19] uses both the previous model and estimated features to compute a semantic correlation between representations during multiple model updates.
The estimated features are used to reproduce the behaviour of older models that are no more available.
推定された特徴は、もはや利用できない古いモデルの振る舞いを再現するために使用される。
0.69
Finally, [18] addresses the problem of lifelong person re-identification in which the previously acquired knowledge is represented as similarity graphs and it is transferred on the current data through graphs convolutions.
Contrastive Supervised Distillation for Continual Representation Learning
連続表現学習のためのコントラスト教師付き蒸留
0.69
3 Reducing feature forgetting with feature distillation is also related to the recent backward compatible representation learning in which newly learned models can be deployed without the need to re-index the existing gallery images [30,31,32].
Contrastive Learning Contrastive learning has been proposed in [35] for metric learning and then it is demonstrated to be effective in unsupervised/self-su pervised representation learning [36,37,21].
In particular, this is achieved as, in the feature space, each image and its augmented samples (the positive samples) are grouped together while the others (the negative samples) are pushed away.
We follow a similar argument which considers as positive also these images.
これらの画像も肯定的と考える類似の議論に従う。
0.64
3 Problem Statement In the continual representation learning problem, a model M(· ; θ, W) is sequentially trained for T tasks on a dataset D = {(xi, yi, ti)| i = 1, 2, . . . , N}, where xi is an image of a class yi ∈ {1, 2, . . . , L}, N is the number of images, and ti ∈ {1, 2, . . . , T} is the task index associated to each image.
3 問題ステートメント 連続表現学習問題において、モデル M(· ; θ, W) はデータセット D = {(xi, yi, ti)| i = 1, 2, . . . . . . . . . . . . . . . L}, N は画像の数、ti ∈ {1, 2, . . . T} は各画像に関連するタスクインデックスである。 訳抜け防止モード: 連続表現学習問題における3つの問題文、モデル m ( · ; θ, ) w ) はデータセット d = { ( xi, xi) 上の t タスクに対して順次訓練される。 yi, ti)| i = 1, 2, . . . , n }, ここで xi はクラス yi ∈ { 1 の像である。 2 . . . . , l }, n は画像の数である。 そして ti ∈ { 1, 2, ... である。 t } は各イメージに関連付けられたタスクインデックスである。
0.86
In particular, for each task k, M is trained on the subset Tk = D|ti=k = {(xi, yi, ti)| ti = k} which represents the k-th training-set that is composed by Lk classes.
特に、各タスク k に対して、m は lk クラスからなる k 番目のトレーニング集合を表す部分集合 tk = d|ti=k = {(xi, yi, ti)| ti = k 上で訓練される。
0.75
Each training-set has different classes and images with respect to the others and only Tk is available to train the model M (memory-free).
At training time of task k, in response to a mini-batch B = {(xi, yi, ti)}|B|
タスク k のトレーニング時には、ミニバッチ B = {(xi, yi, ti)}|B| に対応する。
0.80
i=1 of Tk, the model M extracts the feature vectors and output logits for each image in the batch, i.e., M(xi) = C(φ(xi)), where φ(·, θ) is the representation model which extracts the feature vector fi = φ(xi) and C is the classifier, which projects the feature vector fi in an output vector zi = C(fi).
M(xi) = C(φ(xi)) ここで φ(·, θ) は特徴ベクトル fi = φ(xi) を抽出する表現モデルであり、C は分類器であり、これは出力ベクトル zi = C(fi) に特徴ベクトル fi を投影する。 訳抜け防止モード: モデルMは、Tkのi=1で特徴ベクトルを抽出し、バッチ内の各画像に対する出力ロジットを出力する。 M(xi ) = C(φ(xi ) ) ここで φ ( ·, θ ) は特徴ベクトル fi = φ(xi ) を抽出する表現モデルである。 そして C は分類器で、 特徴ベクトル fi を出力ベクトル zi = C(fi) に投影する。
0.83
At the end of the training phase, M is used to index a gallery-set G = {(xg, yg)| g = 1, 2, . . . , Ng} according to the extracted feature vectors {(fg, yg)}Ng g=1.
トレーニングフェーズの最後に、m は、抽出された特徴ベクトル {(fg, yg)}ng g=1 に従って、ギャラリー集合 g = {(xg, yg)| g = 1, 2, . . . , ng} をインデックスするために使用される。
0.85
At test time, a query-set Q = {xq | q = 1, 2, . . . , Nq} is processed by the representation model φ(·, θ) in order to obtain the set of feature vectors {fq}Nq q=1.
To mitigate the effect of catastrophic forgetting while acquiring novel knowledge from incoming data, we propose a training procedure that follows the teacher-student framework, where the teacher is the model before the update and the student is the model that
Fig. 1: Proposed method is based on the teacher-student framework.
第1図:提案手法は教師・生徒の枠組みに基づく。
0.81
During the training of the student, CE and triplet losses are minimized to learn the new task data, are KD and CSD are used to preserve the old knowledge using the teacher (not trainable).
With reference to Fig 1, at each task k, the student is trained on the training-set Tk = {(xi, yi, ti)| ti = k} and the teacher is set as frozen, i.e., not undergoing learning.
フィグ1について、各タスク k において、生徒はトレーニングセット Tk = {(xi, yi, ti)| ti = k} で訓練され、教師は凍結、すなわち学習を行わないように設定される。
0.70
The loss function that is minimized during the training of the student is the following:
学生の訓練中に最小化される損失関数は以下のとおりである。
0.80
L = Lplasticity + Lstability
L = L可塑性 + Lstability
0.42
(2) where Lstability = 0 during the training of the model on the first task.
(2) 最初のタスクでモデルのトレーニング中に、Lstability = 0となる。
0.59
In the following, the components of the plasticity and stability loss are analyzed in detail.
以下に、可塑性と安定性損失の成分を詳細に分析する。
0.67
In particular, we adopt the following notation.
特に、以下の表記を採用する。
0.55
Given a mini-batch B of training data, both the student and the teacher networks produce a set of feature vectors and classifier outputs in response to training images xi ∈ B. We refer to as {fi}, {zi} for the feature vectors and classifier outputs of the student, respectively, with {f(cid:48) i} for the teacher ones, and with |B| to the number of elements in the mini-batch.
(b) With CSD samples belonging to the same class (same symbol) are clustered together and separated from the others.
b)同じクラスに属するCSDサンプル(サムシンボル)をまとめて、他のものと分離する。
0.67
4.2 Stability Loss The stability loss preserves the previously acquired knowledge in order to limit the catastrophic forgetting effect, that is typically performed using the teacher model for distillation.
The stability loss we propose is formulated as follows:
提案する安定性損失は以下のとおり定式化される。
0.72
Lstability = λKD LKD + λCSD LCSD
Lstability = λKD LKD + λCSD LCSD
0.47
(6) where λKD and λCSD are two weights factors that balance the two loss components, namely Knowledge Distillation (KD) and the proposed Contrastive Supervised Distillation (CSD).
In our experimental results, we set both λKD and λCSD to 1.
実験の結果,λKDとλCSDの両方を1に設定した。
0.76
An evaluation of different values is reported in the ablation studies of Sec. 6.
Sec.6のアブレーション研究では,異なる値の評価が報告されている。
0.75
Knowledge Distillation. KD [39] minimizes the log-likelihood between the classifier outputs of the student and the soft labels produced by the teacher, instead of the groundtruth labels (yi) used in the standard cross-entropy loss.
i j exp(cid:0)zi (cid:1) j=1 exp(cid:0)zj (cid:1) log (cid:80)|B|
私は j j=1 exp(cid:0)zj (cid:1) log (cid:80)|B|
0.46
(cid:1) LKD = 1|B|
(cid:1)。 LKD = 1|B|
0.54
(7) (8) Contrastive Supervised Distillation.
(7) (8) 対照的な監督蒸留
0.45
We propose a new distillation loss, i.e., the Contrastive Supervised Distillation (CSD) that aligns current and previous feature models of the same classes while simultaneously pushing away features of different classes.
the alignment of the student representations to the ones of the same class of the teacher model, which acts as anchors.
生徒の表現と教師モデルの同じクラスの表現のアラインメントは、アンカーとして振る舞う。
0.58
In Fig 2, we show the effect of CSD loss on four samples i=1 with yi ∈ {1, 2}.
図2では、yi ∈ {1, 2} を持つ4つのサンプル i=1 に対するcsd損失の効果を示す。
0.79
Initially (Fig. 2 (a)) the feature vectors extracted by the {(xi, yi)}4 student fi (orange samples) are separated from the teacher ones f(cid:48) i (blue samples).
CSD clusters together features of the same class moving the student representations, which are trainable, towards the fixed ones of the teacher while pushing apart features belonging to different classes.
For the sake of simplicity, this effect is shown just for f(cid:48) 1 and f(cid:48) 1, while f3 and f4 are spaced apart with respect to f(cid:48) 3 which attracts f3 and f4 and push away f1 and f2 as shown in Fig 2
CSD imposes a penalty on feature samples considering not only the overall distribution of features of the teacher model with respect to the student one, but it also clusters together samples of the same class separating from the clusters of the other classes.
Our method differs from KD as the loss function is computed directly on the features and not on the classifier outputs resulting in more discriminative representations.
CSD also considers all the samples of each class as positive samples that are aligned with the same anchor of the teacher and not pairs (teacher-student) of samples as in [40].
5 Experimental Results We perform our experimental evaluation on CIFAR-100 [41] and two fine-grained datasets, namely CUB-200 [42] and Stanford Dogs [43].
We trained the model for 800 epochs for each task using Adam optimizer with a learning rate of 1 · 10−3 for the initial task and 1 · 10−5 for the others.
Random crop and horizontal flip are used as image augmentation.
画像増強にはランダム作物と水平フリップが使用される。
0.72
Following [19], we adopt pretrained Google Inception [47] as representation model architecture on CUB-200 and Stanford Dogs with 512-dimension feature space.
We trained the model for 2300 epochs for each task using with Adam optimizer with a learning rate of 1 · 10−5 for the convolutional layers and 1 · 10−6 for the classifier.
Our approach achieves the highest recall when evaluated on the initial task and the highest recall on the second task between methods trying to preserve old knowledge, being second only to Fine-Tuning that focuses only on learning new data.
This results in our method achieving the highest average recall value with an improvement of ∼2% RECALL@1 with respect to LwF and MMD loss and 10.4% with respect to the FineTuning baseline.
The gap between all the continual representation learning methods and Joint Training is significant (∼8%).
すべての連続的な表現学習方法と合同トレーニングのギャップは重要である(8%)。
0.69
This underlines the challenges of CIFAR-100 in a continual learning scenario since there is a noticeable difference in the appearance between images of different classes causing a higher feature forgetting.
As an upper bound reference, we report the Joint Training performance obtained using all the data to train the model.
上界基準として,全てのデータを用いてモデルを訓練したジョイントトレーニング性能について報告する。
0.77
We report in Tab. 2 the scores obtained with T = 1 on the fine-grained datasets.
タブで報告します。 2 きめ細かいデータセット上で t = 1 で得られるスコア。
0.70
On Stanford Dogs, our approach achieves the highest recall when evaluated on the initial task and comparable result with other methods on the final task with a gap of only 0.2% with respect to Fine-Tuning that focus only on learning new data.
This results in our method achieving the highest average recall value with an improvement of 0.5% RECALL@1 concerning Feature Estimation, 0.8% for MMD loss, and 3.4% for FineTuning.
Our method achieves the highest average recall value with an improvement of 2.1% RECALL@1 with respect to Feature Estimation, 2.9% for MMD loss, and 5.8% for Fine-Tuning.
Differently from CIFAR-100, on fine-grained datasets, there is a lower dataset shift between different tasks leading to a higher performance closer to the Joint Training upper bound due to lower feature forgetting.
(b) the challenging cases of CUB-200 with T = 4 and T = 10, respectively.
(b) T = 4 と T = 10 の CUB-200 の難解な場合。
0.74
These experiments show, consistently with Tab.
これらの実験はTabと一貫して行われた。
0.55
2, how our approach outperforms state-of-the-art methods.
私たちのアプローチは最先端のメソッドよりも優れています。
0.31
In particular, with T = 10 (Fig.
特に、t = 10 の場合(図)。
0.68
4 (b)), our method preserves the performance obtained on the initial task during every update.
4 (b))本手法は,更新毎に初期タスクで得られる性能を保持する。
0.58
CSD largely improves over the state-of-the-art methods by almost 20% - 25% with respect to [19] and [17] achieving similar performance to the Joint Training upper bound.
By leveraging labels information for distillation during model updates, CSD provides better performance and favorably mitigates the catastrophic forgetting of the representation compared to other methods that do not make use of this information.
In Fig 5, we explore the benefits given by the components of the loss in Eq 2 (i.e., CE, triplet, KD, and CSD) and their combinations in terms of RECALL@1 on CUB-200 with T = 10.
When CSD is used, (i.e., CE+CSD, CE+KD+CSD, CE+triplet+CSD, CE+triplet+KD+CSD), we achieve higher RECALL@1 and maintain a more stable trend with respect to others.
Our approach tackles the problem without storing data of previously learned tasks while learning a new incoming task.
提案手法では,新しいタスクを学習しながら,学習したタスクのデータを保存することなく問題に取り組む。
0.69
CSD allows to minimize the discrepancy of new and old features belonging to the same class, while simultaneously pushing apart features from different classes of both current and old data in a contrastive manner.
We evaluate our approach and compare it to state-ofthe-art works performing empirical experiments on three benchmark datasets, namely CIFAR-100, CUB-200, and Stanford Dogs.
Experiments also provide further evidence that feature forgetting evaluated in visual retrieval tasks is not as catastrophic as in classification tasks.
How transferable are features in deep neural networks?
転送可能な機能とは ディープニューラルネットワークで?
0.63
Advances in Neural Information Processing Systems, 2014.
ニューラル情報処理システム(2014年)の進歩
0.78
4. Wei Chen, Yu Liu, Weiping Wang, Erwin Bakker, Theodoros Georgiou, Paul Fieguth, Li Liu, and Michael S Lew.
4.Wei Chen, Yu Liu, Weiping Wang, Erwin Bakker, Theodoros Georgiou, Paul Fieguth, Li Liu, Michael S Lew
0.37
Deep image retrieval: A survey.
深部画像検索:調査。
0.58
arXiv preprint arXiv:2101.11282, 2021.
arxiv プレプリント arxiv:2101.11282, 2021。
0.40
5. Giorgos Tolias, Ronan Sicre, and Herv´e J´egou.
5.ジョルゴス・トリアス、ロナン・シクレ、ヘルヴ・イ・ジェグウ。
0.51
Particular object retrieval with integral maxpooling of cnn activations.
cnnアクティベーションの積分極大化による特に物体の検索
0.72
In ICLR 2016-International Conference on Learning Representations, pages 1–12, 2016.
ICLR 2016-International Conference on Learning Representations, page 1-12, 2016
0.45
6. Joe Yue-Hei Ng, Fan Yang, and Larry S Davis.
6.Joe Yue-Hei Ng、Fan Yang、Larry S Davis
0.34
Exploiting local features from deep networks for image retrieval.
画像検索のためのディープネットワークからのローカル特徴のエクスプロイト。
0.68
In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 53–61, 2015.
IEEEのProceedings of the Proceedings of the IEEE conference on computer vision and pattern recognition Workshops, page 53-61, 2015 訳抜け防止モード: In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops(英語) 53-61頁、2015年。
0.76
7. W Nicholson Price and I Glenn Cohen.
7. ニコルソン・プライスとグレン・コーエン
0.48
Privacy in the age of medical big data.
医療ビッグデータ時代のプライバシー。
0.54
Nature medicine, 25(1):37–43, 2019.
自然 医学、25(1):37-43、2019。
0.52
8. Andrea Cossu, Marta Ziosi, and Vincenzo Lomonaco.
8.アンドレア・コスス、マルタ・シオシ、ヴィンチェンツォ・ロモナコ。
0.63
Sustainable artificial intelligence through continual learning.
持続可能な人工知能 継続的な学習を通じてです
0.55
arXiv preprint arXiv:2111.09437, 2021.
arxiv プレプリント arxiv:2111.09437, 2021。
0.40
9. Michael McCloskey and Neal J Cohen.
9:michael mccloskeyとneal j cohen。
0.60
Catastrophic interference in connectionist networks: In Psychology of learning and motivation, volume 24,
コネクショニストネットワークにおける破滅的な干渉:学習とモチベーションの心理学における第24巻
0.73
The sequential learning problem. pages 109–165.
逐次学習の問題。 109-165頁。
0.55
Elsevier, 1989.
1989年、エルゼヴィエ。
0.61
10. Roger Ratcliff.
ロジャー・ラトクリフ(Roger Ratcliff)。
0.47
Connectionist models of recognition memory: constraints imposed by learn-
認識記憶のコネクショニストモデル--学習によって課される制約-
0.77
ing and forgetting functions.
ing と forgeting 関数。
0.63
Psychological review, 97(2):285, 1990.
心理学的考察,97(2):285,1990
0.80
11. Mochitha Vijayan and SS Sridhar.
11.mochitha vijayanとss sridhar。
0.57
Continual learning for classification problems: A survey.
分類問題に対する継続的な学習:調査
0.81
In International Conference on Computational Intelligence in Data Science, pages 156–166.
データサイエンスにおける計算知能に関する国際会議 (international conference on computational intelligence in data science) 156–166頁。
0.51
Springer, 2021.
2021年。
0.39
英語（論文から抽出）
日本語訳
スコア
Contrastive Supervised Distillation for Continual Representation Learning
連続表現学習のためのコントラスト教師付き蒸留
0.69
11 12. Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars.
11 12.Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, Tinne Tuytelaars
0.39
A continual learning survey: Defying forgetting in classification tasks.
連続学習調査:分類課題における忘れの解消。
0.80
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021。
0.40
13. Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost van de Weijer.
13.Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, Joost van de Weijer 訳抜け防止モード: 13.Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta Andrew D Bagdanov氏とJoost van de Weijer氏。
0.40
Class-incremental learning: survey and performance evaluation on image classification.
クラスインクリメンタルラーニング:画像分類に関する調査と性能評価。
0.82
arXiv preprint arXiv:2010.15277, 2020.
arxiv プレプリント arxiv:2010.15277, 2020
0.43
14. German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter.
Continual lifelong learning with neural networks: A review.
ニューラルネットワークによる持続的生涯学習 : レビュー
0.66
Neural Networks, 113:54–71, 2019.
ニューラルネットワーク, 113:54-71, 2019。
0.65
15. Eden Belouadah, Adrian Popescu, and Ioannis Kanellos.
15) Eden Belouadah, Adrian Popescu, Ioannis Kanellos
0.30
A comprehensive study of class incremental learning algorithms for visual tasks.
授業の総合的研究 視覚タスクのためのインクリメンタル学習アルゴリズムです
0.74
Neural Networks, 135:38–54, 2021.
ニューラルネットワーク, 135:38–54, 2021。
0.67
16. MohammadReza Davari and Eugene Belilovsky.
16. mohammadreza davariとeugene belilovsky。
0.36
Probing representation forgetting in continual learning.
連続学習で忘れる表現の探索。
0.59
In NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications, 2021.
neurips 2021 workshop on distribution shifts: connecting methods and applications, 2021 (英語)
0.39
17. Wei Chen, Yu Liu, Weiping Wang, Tinne Tuytelaars, Erwin M. Bakker, and Michael S. Lew.
17. Wei Chen, Yu Liu, Weiping Wang, Tinne Tuytelaars, Erwin M. Bakker, Michael S. Lew 訳抜け防止モード: 17. wei chen, yu liu, weiping wang, tinne tuytelaars。 アーウィン・m・バクカーとマイケル・s・ルー。
0.56
On the exploration of incremental learning for fine-grained image retrieval.
きめ細かい画像検索のためのインクリメンタル学習の探索について
0.66
In BMVC. BMVA Press, 2020.
BMVC所属。 BMVA、2020年。
0.69
18. Nan Pu, Wei Chen, Yu Liu, Erwin M Bakker, and Michael S Lew.
18.Nan Pu, Wei Chen, Yu Liu, Erwin M Bakker, Michael S Lew
0.34
Lifelong person reidentification via adaptive knowledge accumulation.
適応的知識蓄積による生涯的人物識別
0.66
In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7901–7910, 2021.
The Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page 7901–7910, 2021。 訳抜け防止モード: IEEE / CVF Conference on Computer Vision and Pattern Recognition に参加して 7901-7910頁、2021年。
0.85
19. Wei Chen, Yu Liu, Nan Pu, Weiping Wang, Li Liu, and Michael S Lew.
第19章 陳章,ユリウ,ナンプ,ワイピングワング,リリウ,ミカエル・スルー 訳抜け防止モード: 19. 済陳氏, 融氏, 南周氏, ワイピング・ワン氏, Li LiuとMichael S Lew。
0.77
Feature estimations based correlation distillation for incremental image retrieval.
インクリメンタル画像検索のための特徴推定に基づく相関蒸留
0.78
IEEE Transactions on Multimedia, 2021.
IEEE Transactions on Multimedia, 2021。
0.40
20. Zhizhong Li and Derek Hoiem.
20. zhizhong liとderek hoiem。
0.35
Learning without forgetting. IEEE transactions on pattern
忘れずに学ぶ。 パターン上のIEEEトランザクション
0.67
analysis and machine intelligence, 40(12):2935–2947, 2017.
解析とマシンインテリジェンス, 40(12):2935–2947, 2017
0.85
21. Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton.
Class-incremental learning with pre-allocated fixed classifiers.
固定分類器によるクラス増分学習
0.74
In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6259–6266.
2020年の第25回国際パターン認識会議(ICPR)にて6259-6266頁。
0.74
IEEE, 2021.
IEEE、2021年。
0.81
26. James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al Overcoming catastrophic forgetting in neural networks.
26. James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al Overcoming Superastrophic forgeting in neural network。 訳抜け防止モード: 26. James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan Tiago Ramalho, Agnieszka Grabska - Barwinska, et al ニューラルネットワークにおける破滅的な忘れを克服する。
0.90
Proceedings of the national academy of sciences, 114(13):3521–3526, 2017.
国立科学アカデミー紀要, 114(13)3521-3526, 2017
0.58
27. Gido M Van de Ven and Andreas S Tolias.
27. Gido M Van de VenとAndreas S Tolias
0.34
Three scenarios for continual learning.
連続学習の3つのシナリオ。
0.78
arXiv preprint arXiv:1904.07734, 2019.
arXiv プレプリントarxiv:1904.07734, 2019。
0.45
28. Heechul Jung, Jeongwoo Ju, Minju Jung, and Junmo Kim.
28.ヘチュル・ジュン、ジュンウー・ジュン、ミンジュ・ジュン、キム・ジュンモ。
0.42
Less-forgetting learning in deep
深部における低鍛造学習
0.58
neural networks.
ニューラルネットワーク。
0.65
arXiv preprint arXiv:1607.00122, 2016.
arXiv preprint arXiv:1607.00122, 2016
0.40
29. A. Gretton, AJ.
29. A. Gretton, AJ
0.41
Smola, J. Huang, M. Schmittfull, KM.
スモラ、j.huang、m. schmittfull、km。
0.52
Borgwardt, and B. Sch¨olkopf.
ボルグワードとb・シュ・ソルコップ。
0.50
Co- variate shift and local learning by distribution matching.
共同 分散マッチングによる変分シフトと局所学習。
0.63
MIT Press, 2009.
2009年、MIT出版。
0.63
英語（論文から抽出）
日本語訳
スコア
12 T. Barletti et al
12 T. Barletti et al
0.46
30. Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto.
30.ヤンタオ・シェン、ユアンジュン・シオン、ウイ・シア、ステファノ・ソアトー
0.55
Towards backward-compatible representation learning.
後方互換性のある表現学習を目指す。
0.47
In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6368–6377, 2020.
In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page 6368–6377, 2020。 訳抜け防止モード: IEEE / CVF Conference on Computer Vision and Pattern Recognition に参加して 6368-6377、2020年。
0.82
31. Federico Pernici, Matteo Bruni, Claudio Baecchi, and Alberto Del Bimbo.
31. Federico Pernici, Matteo Bruni, Claudio Baecchi, Alberto Del Bimbo
0.35
Regular polytope networks.
通常のポリトープ ネットワーク。
0.64
IEEE Transactions on Neural Networks and Learning Systems, 2021.
IEEE Transactions on Neural Networks and Learning Systems, 2021。
0.41
32. Niccol´o Biondi, Federico Pernici, Matteo Bruni, and Alberto Del Bimbo.
32. Niccol ́o Biondi, Federico Pernici, Matteo Bruni, Alberto Del Bimbo
0.39
Cores: Compatible
Cores: 互換性
0.83
representations via stationarity, 2021.
静止線による表現 2021年
0.64
33. Rahaf Aljundi, Klaas Kelchtermans, and Tinne Tuytelaars.
In Proceedings of the IEEE/CVF CVPR, pages 11254–11263, 2019.
院 IEEE/CVF CVPR, page 11254–11263, 2019。
0.42
34. Federico Pernici, Matteo Bruni, and Alberto Del Bimbo.
34. フェデリコ・ペルニッチ、マテオ・ブルニ、アルベルト・デル・ビンボ
0.49
Self-supervised on-line cumulative learning from video streams.
ビデオストリームからの自己教師付きオンライン累積学習
0.61
Computer Vision and Image Understanding, 197:102983, 2020.
コンピュータビジョンとイメージ理解 197:102983, 2020。
0.84
35. Sumit Chopra, Raia Hadsell, and Yann LeCun.
35.Sumt Chopra,Raia Hadsell,Yann LeCun
0.27
Learning a similarity metric discriminatively, with application to face verification.
類似度メトリックを識別的に学習し、顔認証に応用する。
0.65
In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05), volume 1, pages 539–546.
2005年、IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05)、1巻539-546頁。 訳抜け防止モード: 2005年IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05)に参加。 巻1、539-546頁。
0.88
IEEE, 2005.
2005年、IEEE。
0.70
36. Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick.
36.カイミン・ヘ、ハチ・ファン、ユキシン・ウー、サイニン・シー、ロス・ガーシック
0.55
Momentum contrast for unsupervised visual representation learning.
教師なし視覚表現学習におけるモメンタムコントラスト
0.62
In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729–9738, 2020.
In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page 9729–9738, 2020。 訳抜け防止モード: IEEE / CVF Conference on Computer Vision and Pattern Recognition に参加して 9729-9738、2020年。
0.82
37. Ishan Misra and Laurens van der Maaten.
37. イシャン・ミスラとローレンス・ファン・デル・マタン
0.55
Self-supervised learning of pretext-invariant representations.
プレテキスト不変表現の自己教師付き学習
0.48
In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707–6717, 2020.
In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page 6707–6717, 2020。 訳抜け防止モード: IEEE / CVF Conference on Computer Vision and Pattern Recognition に参加して 6707-6717頁、2020年。
0.85
38. Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan.
38. Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, Dilip Krishnan 訳抜け防止モード: 38 . Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu そして、ディリップ・クリシュナン。
0.73
Supervised contrastive learning.
対照的な学習を監督する。
0.43
In NeurIPS, 2020.
2020年、NeurIPS。
0.70
39. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean.
39 ジョフリー・ヒントン、オリオール・ヴィニールズ、ジェフ・ディーン
0.46
Distilling the knowledge in a neural network.
ニューラルネットワークで知識を蒸留する。
0.66
arXiv preprint arXiv:1503.02531, 2015.
arXiv preprint arXiv:1503.02531, 2015
0.40
40. Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio.
40.Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, Yoshua Bengio 訳抜け防止モード: 40. Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou アントワーヌ・チャサン、カルロ・ガッタ、ヨシュア・ベンジオ。
0.81
Fitnets: Hints for thin deep nets.
Fitnets: 薄いディープネット用のヒント。
0.72
arXiv preprint arXiv:1412.6550, 2014.
arxiv プレプリント arxiv:1412.6550, 2014
0.41
41. Alex Krizhevsky, Geoffrey Hinton, et al Learning multiple layers of features from tiny
41. alex krizhevsky, geoffrey hinton, et al ミニチュアから複数の特徴の層を学ぶ
0.79
images. 2009. 47. Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich.
画像。 2009. 47.Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich
0.50
Going deeper with convolutions.
畳み込みでさらに深く進む。
0.61
In Proceedings of the IEEE conference CVPR, 2015.
2015年、ieee conference cvprで開催。
0.51
48. Herve Jegou, Matthijs Douze, and Cordelia Schmid.
48 ハーヴェ・ジェゴウ、マティジス・ドゥーゼ、コーデリア・シュミード
0.37
Product quantization for nearest neighbor search.
近接探索のための製品定量化
0.55
IEEE transactions on pattern analysis and machine intelligence, 33(1):117–128, 2010.
Novel dataset for fine-grained image categorization.
きめ細かい画像分類のための新しいデータセット
0.67
In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, June 2011.
The First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, June 2011 訳抜け防止モード: 第1回ファイン・グラインド視覚分類ワークショップ, IEEE Conference on Computer Vision and Pattern Recognition コロラドスプリングス、CO、2011年6月。
Deep metric learning via lifted structured feature embedding.
昇降型特徴埋め込みによる深度測定学習
0.55
In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004–4012, 2016.
Proceedings of the IEEE conference on computer vision and pattern recognition, page 4004–4012, 2016 訳抜け防止モード: In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 4004-4012頁、2016年。
0.83
45. Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R Scott.
45.Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, Matthew R Scott
0.35
Multisimilarity loss with general pair weighting for deep metric learning.
ディープメトリック学習のための一般対重み付き多相性損失
0.82
In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019.
コンピュータビジョンとパターン認識に関するieee/cvfカンファレンスの議事録、2019年。
0.71
46. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
46.開明,Xiangyu Zhang,Shaoqing Ren,Jian Sun
0.26
Deep residual learning for imIn Proceedings of the IEEE conference on computer vision and pattern
imInの深層学習 : コンピュータビジョンとパターンに関するIEEEカンファレンスの成果
0.64
age recognition. recognition, pages 770–778, 2016.