CoChat
Sign Up
Cited by 90Domain Adaptation and Few-Shot LearningAdvanced Neural Network ApplicationsAdvanced Image and Video Retrieval Techniques

IncDet: In Defense of Elastic Weight Consolidation for Incremental Object Detection

Open access
Liu, Liyang, Kuang, Zhanghui, Chen, Yimin, Xue, Jing-Hao, Yang, Wenming, Zhang, Wayne·2021·IEEE Transactions on Neural Networks and Learning Systems
DOI 10.1109/tnnls.2020.3002583
Open this paper in CoChat — free accountOpen PDF

Abstract

Elastic weight consolidation (EWC) has been successfully applied for general incremental learning to overcome the catastrophic forgetting issue. It adaptively constrains each parameter of the new model not to deviate much from its counterpart in the old model during fine-tuning on new class data sets, according to its importance weight for old tasks. However, the previous study demonstrates that it still suffers from catastrophic forgetting when directly used in object detection. In this article, we show EWC is effective for incremental object detection if with critical adaptations. First, we conduct controlled experiments to identify two core issues why EWC fails if trivially applied to incremental detection: 1) the absence of old class annotations in new class images makes EWC misclassify objects of old classes in these images as background and 2) the quadratic regularization loss in EWC easily leads to gradient explosion when balancing old and new classes. Then, based on the abovementioned findings, we propose the corresponding solutions to tackle these issues: 1) utilize pseudobounding box annotations of old classes on new data sets to compensate for the absence of old class annotations and 2) adopt a novel Huber regularization instead of the original quadratic loss to prevent from unstable training. Finally, we propose a general EWC-based incremental object detection framework and implement it under both Fast R-CNN and Faster R-CNN, showing its flexibility and versatility. In terms of either the final performance or the performance drop with respect to the upper bound of joint training on all seen classes, evaluations on the PASCAL VOC and COCO data sets show that our method achieves a new state of the art.

Integrity signals

Registry

Crossref · 10.1109/tnnls.2020.3002583

Cites

12 of 12 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 28 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Certain parameters have extremely large gradients before gradient explosion occurs.

    “we notice that certain parameters own extremely large gradients before gradient explosion.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.
  • Findingown data

    The quadratic regularization loss in EWC tends to cause gradient explosion when balancing old and new classes.

    “the quadratic regularization loss in EWC easily leads to gradient explosion when balancing old and new classes”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.
  • Findingown data

    When EWC is applied to incremental detection without adaptation, the lack of old-class annotations in new-class images causes old-class objects to be misclassified as background.

    “the absence of old class annotations in new class images makes EWC misclassify objects of old classes in these images as background”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.
  • Findingown data

    Huber regularization significantly outperforms traditional gradient clipping.

    “It also differs from and significantly outperforms the traditional gradient clipping method [17] (Fig. 5).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [17] source

      Not checked in the free run.

    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.
  • Findingown data

    Adding several classes simultaneously produces higher mAP than adding a single class in each step.

    “Besides, comparing Tab.VI and Tab.VIII, we see that adding multiple classes simultaneously manifests better mAP than adding a single class each time.”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim. 7
  • May combine two claims. 44
  • The quote only partly states this claim. 2

  • Classified as probably not a substantial claim. May combine two claims.

    Findingreasoning

    Huber regularization is smooth and equals the quadratic loss when the weighted parameter deviation is within the clipping threshold.

    “as the Huber regularization is smooth so it is the same as the quadratic loss when p w |✓ ✓ ?| 6 .”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Findingown data

    In pilot experiments on VOC 2007 with Fast R-CNN, masking old-class objects in new images raised mAP of the 15 old classes from 46.9% to 59.7% and mAP of all 20 classes from 47.6% to 56.7%.

    “Specifically, mAP of old classes increases from 46.9% to 59.7% while that of all classes boosts from 47.6% to 56.7%.”
    Quote in the paperSubstantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    On PASCAL VOC and COCO, IncDet achieves new state-of-the-art results in final performance and in performance drop relative to the joint-training upper bound.

    “evaluations on the PASCAL VOC and COCO datasets show that our method achieves a new state-of-the-art.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Findingown data

    Increasing the balancing parameter λ improves old-class performance significantly while new-class performance does not change notably.

    “larger leads to significantly better performance on old classes but that of new classes does not change notably.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Findingown data

    When λ reaches a certain value, such as 10, training becomes dramatically unstable, preventing identification of the best old-new trade-off.

    “when reaches a certain value (e.g.10), the training process becomes dramatically unstable, which prevents us from finding the best trade-off between old and new classes.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Findingown data

    Gradient explosion still occurs when oracle manual labels of old classes are used in new images, so inaccurate pseudo annotations are not the cause.

    “the gradient explosion still happens even if we use the oracle manual labels of old classes in the new images.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Findingown data

    Fine-tuning with pseudo annotations corrected by new-class manual labels raised mAP of the 15 old classes to 70.6% and mAP of all 20 classes to 66.3%, which is superior to the masking strategy.

    “mAP of the 15 old classes and that of all classes increases to 70.6% and 66.3%, respectively, which is remarkably superior to its counterpart with masking training strategy”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    mAP decreases as the number of false detections increases, but performance is fairly robust to the exact detection score threshold.

    “the mAP decreases as the number of false detections increases but the performance is fairly robust to the exact setting of .”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Against auxiliary-based methods using proposal-based or proposal-free frameworks, IncDet achieves much higher new-class mAP and comparable or higher old-class mAP.

    “Compared with the auxiliary-based methods implemented with either proposalbased [8] or proposal-free framework [21], [22], our approach performs substantially better regarding mAP for new classes and meanwhile obtains comparable or better mAP for old ones, indicating we can leverage the capacity of the network to a greater extent.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [8] source

      Not checked in the free run.

    • cites [21] source

      Not checked in the free run.

    • cites [22] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    The results in Table IV are lower than those in Figure 6 because the validation set is excluded, so less training data is used.

    “Note that the results in Tab.IV are lower than those in Fig. 6 because less training data is used (the validation set is excluded).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    IncDet surpasses previous best methods in both final performance and the performance gap compared with the joint-training upper bound.

    “our results surpass the previous best [8], [21], [22], [23] in terms of both the final performance and the performance gap compared with the joint training upper bound.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [8] source

      Not checked in the free run.

    • cites [21] source

      Not checked in the free run.

    • cites [22] source

      Not checked in the free run.

    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Findingown data

    Using Huber regularization instead of quadratic loss, with or without gradient clipping, gives a better balance between old and new classes and a higher overall mAP.

    “Compared with the quadratic loss both with and without gradient clipping [17], we are able to obtain better old and new class trade-off and thus superior final mAP across all learned classes.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [17] source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Relative to ILOD, IncDet shows a similar drop in mAP@0.5 but a much smaller drop in average mAP over IoU thresholds from 0.5 to 0.95.

    “Compared with ILOD [8], our method achieves comparable performance drop in terms of mAP@0.5, and much less performance drop in terms of average mAP across different IoU thresholds (0.5 to 0.95).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [8] source

      Not checked in the free run.

  • May combine two claims.

    Findingnone given

    Fine-tuning remains stable even with a very large lambda, such as 1000, and no gradient explosion occurs.

    “We can even fine-tune the model stably with a huge such as 1000 without gradient explosion.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    When instantiated with Fast R-CNN or Faster R-CNN, IncDet consistently beats the knowledge-distillation-based method in both final performance and drop from upper bounds.

    “They show that our method instantiated with Fast R-CNN and Faster R-CNN outperforms the approach based on knowledge distillation [53] consistently, in terms of both the final performance and the performance drop compared with the upper bounds.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [8] source

      Not checked in the free run.

    • cites [53] source

      Not checked in the free run.

  • May combine two claims.

    Findingreasoning

    Both IncDet and ILOD retain old-class performance well but have difficulty detecting new classes.

    “Both methods perform well on memorizing old classes but struggle to detect new classes.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    At test time, adding five classes increases time by roughly one millisecond and memory by roughly 0.2 MB.

    “As can be seen from Tab. XI adding 5 classes only spends about 1 more millisecond and consumes about 0.2 MB more memory.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The best result from gradient clipping is 67.1% mAP, which is below the 69.0% achieved by Huber regularization, demonstrating that the adaptive clipping method is superior.

    “Although gradient clipping can alleviate the gradient explosion issue, its best performance 67.1% lags behind that of our Huber regularization 69.0%, showing the superiority of our adaptive clipping method.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [17] source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    In the group-wise incremental setting, IncDet substantially outperforms CIFRCN on per-group mAP, overall mAP, and drop from joint-training upper bounds.

    “The results are shown in Tab.VII, where our method significantly outperforms its counterpart in terms of the mAP for each group, the mAP for all classes and the mAP drop compared with joint training upper bounds, showing that our method is also superior in the group-wise incremental detection setting.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [23]

      Not checked in the free run.

  • May combine two claims.

    Findingreasoning

    When IncDet is incrementally fine-tuned on COCO, performance is better than fine-tuning on VOC and even exceeds joint training on VOC with all seen classes.

    “Tab.X shows that the performance obtained via incrementally fine-tuning on COCO with our method is better than that of fine-tuning on VOC, even better than that of jointly-training on VOC with all seen classes.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Huber regularization is crucial to the proposed approach; the network cannot be trained successfully without it.

    “The Huber regularization is extremely important in our approach, without which we cannot train the network successfully.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Increasing lambda raises mAP for old classes and lowers mAP for new classes.

    “Generally speaking, as increases, mAP of the old classes increases and that of the new classes decreases (shown in Tab.V (a)).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In both in-domain and cross-domain settings, IncDet shows a much smaller performance drop than ILOD.

    “Compared with ILOD [8], our IncDet achieves much less performance drop in both in-dataset and cross-dataset scenarios.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [8] source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    IncDet is substantially more efficient and economical in both computation and memory.

    “Tab.XII verifies that our IncDet is much more efficient and economic in terms of computation and memory.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    When implemented with Fast R-CNN or Faster R-CNN, IncDet surpasses prior state-of-the-art methods while requiring less memory and computation during incremental fine-tuning.

    “Comprehensive experiments show that our general approach implemented with either Fast R-CNN or Faster R-CNN outperforms the previous best methods [8], [21], [22], [23] while requiring less memory and computation during incremental fine-tuning.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [8] source

      Not checked in the free run.

    • cites [21] source

      Not checked in the free run.

    • cites [22] source

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    The Faster R-CNN version of IncDet consumes more time and memory than the Fast R-CNN version because it includes an RPN module, yet it remains more efficient than ILOD.

    “IncDet based on Faster R-CNN needs more time and memory compared with the Fast R-CNN variant because it includes an RPN module for online proposal generation, but it is still more efficient than the counterpart ILOD [8].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [8] source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Controlled experiments identified the main problems that prevent direct application of EWC to object detection, and these were solved with pseudo annotation and Huber regularization.

    “Key issues causing the failure of directly applying EWC to object detection are identified by controlled experiments and addressed by our simple but effective pseudo annotation and Huber regularization.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    IncDet is a general EWC-based incremental object detection framework that can be implemented under both Fast R-CNN and Faster R-CNN, showing flexibility and versatility.

    “we propose a general EWC-based incremental object detection framework and implement it under both Fast R-CNN and Faster R-CNN, showing its flexibility and versatility.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Interpretationreasoning

    Distributional shifts across classes cause model weights to change drastically at the start of incremental fine-tuning, making gradients grow and explode.

    “Due to the distributional shifts of the images across different classes, the model weights will change drastically to accommodate the new classes at the beginning of incremental fine-tuning, and thus the gradients are very easy to grow dramatically and finally explode.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Interpretationreasoning

    New classes can be added to IncDet sequentially or all at once.

    “So the new classes can be seamlessly added sequentially or all at once.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Interpretationreasoning

    Because IncDet is versatile, it can be combined with advanced detectors such as Faster R-CNN, leading to much better final performance on difficult datasets.

    “Also, thanks to the versatility of our IncDet, we can leverage advanced detection methods such as Faster R-CNN and thus achieve much better final performance in challenging datasets.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Compared with conventional holistic gradient clipping, the proposed adaptive gradient clipping both stabilizes training and achieves better performance.

    “Compared with the conventional holistic gradient clipping [17], the proposed adaptive gradient clipping not only realizes stable training but also achieves better performance.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [17] source

      Not checked in the free run.

    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The superior result from COCO fine-tuning is attributed to COCO containing greater appearance variation in scale, pose, and viewpoint than VOC.

    “We believe that this is because COCO has much more appearance variations such as scale, pose and viewpoint than VOC.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    If lambda is too large, the gradient is not constrained effectively, which leads to worse performance on new classes.

    “However when is too large, the gradient cannot be constrained effectively and thus leads to worse performance on new classes.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    A possible reason for the difficulty in detecting new classes is overfitting to the new-class images, because only a few hundred images are added per class.

    “Overfitting on new category images can be the reason, since only hundreds of images are added for each class.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based). May combine two claims. The quote only partly states this claim.

    Interpretationreasoning

    When λ = +∞, the new model remains the same as the old one, so no new classes can be learned.

    “Ideally, when = +1, the new model is kept the same as the old one, and thus no new classes can be learned.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Interpretationreasoning

    Pseudo annotations of old classes prevent the model from misclassifying old-class objects as new classes, reducing false alarms of new classes during testing.

    “We believe this is because pseudo annotations of the old classes can prevent the model from misclassifying old class objects as the new classes, and hence reduce false alarms of the new classes during testing.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The regularized objective depends on both the likelihood of new data and the weighted deviation of parameters from the old model, with weights set by each parameter's importance for all seen tasks.

    “Intuitively, the regularized objective not only depends on the likelihood of the new data, but also the weighted model parameter deviation from the old model, with the weight decided by the importance of each parameter for all seen tasks.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Interpretationreasoning

    If lambda is too small, the regularization loss approaches zero, so incremental fine-tuning behaves like naive fine-tuning and the model forgets old classes.

    “When is too small, the incremental fine-tuning degrades to the naïve fine-tuning without regularization as the regularization loss approaches zero, so the model will forget the old classes.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    In incremental detection, the model may predict new classes in new images as old classes with high confidence when appearances are similar, so pseudo ground truth must be generated carefully.

    “in incremental detection, the model may predict new classes in new images as old classes with high confidence if they share similar appearances, which should be dealt with carefully to generate accurate pseudo ground truth.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    IncDet is more general than ILOD and can be adopted in any conventional object detection method, not only those depending on external proposal generation.

    “IncDet is more general and can be adopted in any conventional object detection method, not only those who depend on external proposal generation approaches.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Interpretationreasoning

    IncDet is auxiliary-free and more efficient and lightweight than previous incremental detection methods because it embeds old-class knowledge in parameter importance and requires only a single network fine-tuning.

    “our auxiliary-free EWC-based IncDet embeds knowledge with parameter importance for old classes.When learning new classes IncDet only demands a single network fine-tuning, which is more efficient and lightweight.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Interpretationown data

    IncDet demonstrates generalization ability for both in-domain and cross-domain settings.

    “This reveals the generalization ability of our method both for the in-domain and cross-domain setting.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    During incremental fine-tuning, the regularization term pushes old-class regions toward foreground while the likelihood term pushes them toward background, causing optimization difficulty and poor convergence.

    “the regularization term in Eq. ( 10) suggests to classify old class regions as foreground, but the likelihood term forces the network to regard them as background.This contradiction leads to the optimization difficulty and finally makes the network struggle to converge well.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    IncDet's flexibility comes from its Bayesian online learning nature.

    “Our framework is flexible thanks to its Bayesian online learning nature.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Study design and conductnone given

    The clipping threshold g in Eq. (15) was tuned so that the quadratic loss matched the Huber regularization loss and gradient explosion was eliminated.

    “Note that the clipping threshold g in Eq. ( 15) is adjusted properly so that the quadratic loss matches the Huber regularization loss and the gradient explosion vanishes.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    When λ = 0, optimization degrades to conventional fine-tuning and leads to catastrophic forgetting.

    “When = 0, the optimization of the model degrades to conventional fine-tuning, and thus leads to catastrophic forgetting.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Data samples are independent and identically distributed.
    • The authors state this as an assumption: The Hessian matrix in EWC has a diagonal structure.

  • May combine two claims.

    Premisecitation

    ILOD relies on an auxiliary teacher network during fine-tuning and therefore must maintain a frozen copy of the entire model in memory, whereas IncDet does not.

    “ILOD exploits one auxiliary teacher network during fine-tuning and thus needs to maintain a frozen copy of the whole model in memory while we do not.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [8] source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Since the extra epoch for importance update is like a normal epoch, the time for importance update is less than one tenth of typical training time, which usually requires more than ten epochs.

    “The extra epoch for importance update is the same as normal epochs for incremental finetuning, so the time spent for importance update is less than one tenth of the usual training time where more than ten epochs are typically needed.”
    Quote in the paperSubstantial claim
49 of 49

  • The authors state this as an assumption.

    Data samples are independent and identically distributed.

    “Under the basic assumption that data samples are independent and identically distributed”

  • The authors state this as an assumption.

    The Hessian matrix in EWC has a diagonal structure.

    “EWC assumes the diagonal structure [11], [15] of the Hessian matrix”
2 of 2

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Response-based knowledge distillation has not been explored in IOD.

    “response-based method [12] has not been explored in IOD [25] yet.”

    From Overcoming Catastrophic Forgetting in Incremental Object Detection via Elastic Response Distillation (2022) · cites this paper as [25]

    Section D, “Incremental object detection” (page not specified in our copy), describes prior IOD methods that distill knowledge by minimizing discrepancies between teacher and student outputs. That contradicts the claim that response-based knowledge distillation had not been explored in IOD. (LLM, conf=0.87)

    Checked by a reasoning model — may be wrong.

    Experimental — accuracy not yet measured

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Last-group results contradict the claim of group-wise superiority

    Major

    For the final five-class group in Table VII, CIFRCN scores 59.6% mAP, while IncDet-Fast and IncDet-Faster score 57.0% and 56.6%. Their drops relative to joint training are also larger for that group, so IncDet does not outperform its comparator on every group.

  • The same ILOD result has two different new-class drops

    Minor

    Tables VIII and X repeat the ILOD results for adding classes 11-20, including the same bounds and measured mAPs, but give the new-class drop as 8.1 and 8.2 percentage points, respectively.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 49

  1. 10.1109/TPAMI.2017.2769047 · DOI
  2. 10.1109/TNNLS.2015.2416353 · DOI
  3. 10.1109/TNNLS.2017.2775225 · DOI
  4. 10.1109/TKDE.2012.136 · DOI
  5. 10.1109/TNN.2011.2160459 · DOI
  6. he, 2011, Incremental learning from stream data, IEEE Trans Neural Netw, 22, 1901, 10.1109/TNN.2011.2171713 · DOI
  7. 10.1109/TNNLS.2018.2866622 · DOI
  8. yoon, 2018, Lifelong learning with dynamically expandable networks, Proc Int Conf Learn Representations, 1
  9. rusu, 2016, Progressive neural networks, arXiv 1606 04671
  10. goyal, 2017, Accurate, large minibatch SGD: Training ImageNet in 1 hour, arXiv 1706 02677
  11. 10.1109/ICCV.2019.00851 · DOI
  12. farquhar, 2018, Towards robust evaluations of continual learning, Proc Privacy Mach Learn Artif Intell Workshop (ICML), 1
  13. schaul, 2018, The barbados 2018 list of open issues in continual learning, arXiv 1811 07004
  14. 10.1016/0304-3975(85)90224-5 · DOI
  15. 10.1109/CVPR.2018.00762 · DOI
  16. 10.1109/ICCV.2017.322 · DOI · Paper page
  17. 10.1007/s11263-009-0275-4 · DOI · Paper page
  18. 10.1109/WACV45572.2020.9093365 · DOI
  19. 10.1109/72.501732 · DOI · Paper page
  20. 10.1109/ICME.2019.00009 · DOI · Paper page
  21. gepperth, 2016, Incremental learning algorithms and applications, Proc Eur Sympoisum Artif Neural Netw (ESANN), 357
  22. cauwenberghs, 2001, Incremental and decremental support vector machine learning, Proc Adv Neural Inf Process Syst, 409
  23. liu, 2016, SSD: Single shot multibox detector, Proc Eur Conf Comput Vis, 21
  24. 10.1109/ICCV.2017.324 · DOI · Paper page
  25. paszke, 2019, Pytorch: An imperative style, high-performance deep learning library, Proc Adv Neural Inf Process Syst, 8024
  26. zitnick, 2014, Edge boxes: Locating object proposals from edges, Proc Eur Conf Comput Vis, 391
  27. ioffe, 2015, Batch normalization: Accelerating deep network training by reducing internal covariate shift, Proc 32nd Int Conf Mach Learn, 37, 448
  28. 10.1109/CVPR.2018.00433 · DOI
  29. 10.1109/ACVMOT.2005.107 · DOI · Paper page
  30. hinton, 2015, Distilling the knowledge in a neural network, ArXiv 1503 02531
  31. 10.1109/TPAMI.2017.2773081 · DOI · Paper page
  32. 10.1109/TNN.2009.2024147 · DOI
  33. zenke, 2017, Continual learning through synaptic intelligence, Proc 34th Int Conf Mach Learning (JMLR), 70, 3987
  34. shin, 2017, Continual learning with deep generative replay, Proc Adv Neural Inf Process Syst, 2990
  35. nguyen, 2018, Variational continual learning, Proc Int Conf Learn Representations, 1
  36. ritter, 2018, Online structured Laplace approximations for overcoming catastrophic forgetting, Proc Adv Neural Inf Process Syst, 3742
  37. opper, 1998, A Bayesian approach to on-line learning, On-Line Learning in Neural Networks, 363
  38. pascanu, 2013, On the difficulty of training recurrent neural networks, Proc Int Conf Mach Learn, 1310
  39. 10.1109/ICCV.2015.169 · DOI · Paper page
  40. 10.1109/CVPR.2009.5206848 · DOI · Paper page
  41. 10.1016/j.neucom.2016.02.011 · DOI
  42. kuznetsova, 2018, The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale, arXiv 1811 00982
  43. lin, 2014, Microsoft COCO: Common objects in context, Proc Eur Conf Comput Vis, 740
  44. 10.1007/s11263-013-0620-5 · DOI · Paper page
  45. 10.1109/CVPR.2014.81 · DOI · Paper page
  46. 10.1109/TPAMI.2016.2601099 · DOI · Paper page
  47. 10.1109/TPAMI.2015.2389824 · DOI · Paper page
  48. 10.1162/neco.1992.4.3.448 · DOI · Paper page
  49. 10.1109/TNNLS.2011.2178124 · DOI

Cited by

Sorted by citations

Authors
Liu, L., Kuang, Z., Chen, Y., Xue, J.-H., Yang, W., & Zhang, W.
Affiliations
  • Tsinghua University (CN)
  • Sensetime (China) (CN)
  • University College London (GB)
ORCID
Venue
IEEE Transactions on Neural Networks and Learning Systems
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
2021-06
Volume
32
Issue
6
Pages
2306-2319
ISSN
  • 2162-237X (ISSN-L)
  • 2162-2388
DOI
10.1109/tnnls.2020.3002583
OpenAlex
W3038556141
Cited by
90
References
12
Citations by year
2026: 18 · 2025: 12 · 2024: 19 · 2023: 19 · 2022: 13 · 2021: 8 · 2020: 1
FWCI
4.03
Type
journal article
Language
en
Open access
Licence
Locations
  • IEEE Transactions on Neural Networks and Learning Systems (publishedVersion)
  • PubMed (publishedVersion)
  • UCL Discovery (University College London) (submittedVersion)
Funders
  • National Key Research and Development Program of China (2016YFB0101001)
  • Natural Science Foundation of Guangdong Province (2020A1515010711)
  • Special Foundation for the Development of Strategic Emerging Industries of Shenzhen (JCYJ20170817161845824)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Image and Video Retrieval Techniques
Keywords
elastic weight consolidation, catastrophic forgetting, incremental learning, gradient explosion, Fast R-CNN, Faster R-CNN, class-incremental learning

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-05.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-05.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review