An Appraisal of Incremental Learning Methods

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The initial survey obtained 353 works; after removing repeated ones, 209 works were selected; after analysis and selection, 109 articles were identified as the target in this review.

“In the initial survey, 353 works were obtained. After identification and removal of the repeated ones, 209 works were selected for the next step. After analysis and selection, 109 articles were identified as the target in this review.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Although experimental protocols differ, dataset selection is often close and general; MNIST, CIFAR-100, and ImageNet ILSVRC are frequently used.

“Although the experimental protocol of each reference is not exactly the same, incremental learning is the basic research to realize artificial intelligence, so the selection of the datasets is often close and general. Some general datasets are frequently used, such as MNIST, CIFAR-100, and ImageNet ILSVRC.”

Classified as probably not a substantial claim.

Researchers are increasingly inclined to use a combination of various existing strategies to study incremental learning.

“Researchers are increasingly inclined to use a combination of various existing strategies to study incremental learning”

May combine two claims.

In terms of strategies to deal with catastrophic forgetting, regularization strategy has received widespread attention, architectural strategy research tends to be stable, and rehearsal and pseudo-rehearsal strategies research is gradually increasing.

“In terms of various strategies to deal with CF, it can be seen from Figure 5b that the regularization strategy has received widespread attention at present, the research on the architectural strategy tends to be stable, and the research on rehearsal and pseudo-rehearsal strategies is gradually increasing.”

Probably not a substantial claim: may not pass S2 (checkable).

The strategy for dealing with catastrophic forgetting has reached a bottleneck.

“Through analysis, we found that the strategy for dealing with catastrophic forgetting has reached a bottleneck.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

The van de Ven et al. framework does not make a good conceptual distinction between incremental learning scenarios.

“However, it does not make a good conceptual distinction between incremental learning scenarios.”

• cites [20] Van de Ven, G.M., and Tolias, A.S. (2019). Three scenarios for continual learning. arXiv.Not checked yet

  Not checked in the free run.

May combine two claims.

To overcome catastrophic forgetting, a model must be both plastic to acquire new knowledge and stable to consolidate existing knowledge, but meeting these two requirements at the same time is very hard.

“To overcome CF, a model must be both plastic to acquire new knowledge and stable to consolidate existing knowledge, but meeting these two requirements at the same time is very hard.”

• AssumptionRests on an assumption: Incremental learning remains a long term challenge.

May combine two claims.

MAS is partly similar to SI but supports the use of unlabeled data to obtain weight sensitivity.

“The MAS method was proposed by Aljundi et al. [49], which was partly similar with SI but it supports the use of unlabeled data to obtain weight sensitivity.”

• cites [49] Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., and Tuytelaars, T. (2018, January 8–14). Memory Aware Synapses: Learning What (not) to Forget. Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany. sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

The keywords incremental learning, lifelong learning and continuous learning were used as input for search engines of Web of Science, IEEEXplore, and DBLP.

“The keywords, incremental learning, lifelong learning and continuous learning were used as input for search engines of Web of Science, IEEEXplore, and DBLP.”

May combine two claims.

The hierarchical tree network method mitigates catastrophic forgetting to a certain extent but consumes more space and is not efficient in training.

“Although this method mitigates catastrophic forgetting to a certain extent, it is intuitively a method that consumes more space and is not efficient in training.”

• cites [28] Roy, 2020, Tree-CNN: A hierarchical Deep Convolutional Neural Network for incremental learning, Neural Netw. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

RWalk improves the regularization methods of EWC and SI and further improves performance, but it still retains some old data for performance reasons, which shows the effectiveness of rehearsal strategies for mitigating catastrophic forgetting.

“Rwalk improves the regularization methods of EWC and SI, and further improves the performance based on them. However, for performance reasons, some old data are still retained. This shows the effectiveness of rehearsal strategies for mitigating catastrophic forgetting.”

• cites [47] Chaudhry, A., Dokania, P.K., Ajanthan, T., and Torr, P.H.S. (2018, January 8–14). Riemannian Walk for Incremental Learning: Understanding Forgetting and Intransigence. Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany. sourceNot checked yet

  Not checked in the free run.

• cites [48] Kullback, 1951, On information and sufficiency, Ann. Math. Stat. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Regularization strategy mitigates catastrophic forgetting by adding a special regularization term to the loss function, with the core idea of limiting parameter updates to improve model stability.

“This strategy mitigates catastrophic forgetting by adding a special regularization term to loss function. The core idea is to limit the updating of parameters to improve the model stability, and thereby alleviate catastrophic forgetting.”

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

van de Ven et al. classified incremental learning scenarios into task incremental, domain incremental and class-incremental scenarios according to specific experimental protocols.

“van de Ven et al. [20] have provided a comprehensive framework for the scenarios of incremental learning; they classified incremental learning scenarios into task incremental scenario, domain incremental scenario and class-incremental scenario according to specific experimental protocols.”

• cites [20] Van de Ven, G.M., and Tolias, A.S. (2019). Three scenarios for continual learning. arXiv.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The three solution strategies-architectural, regularization, and rehearsal/pseudo-rehearsal-do not contradict each other but cooperate with each other.

“It is worth emphasizing that these three strategies do not contradict with each other, instead, they cooperate with each other.”

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

A non-iterative learning method is highly efficient because it is a globally ordered and locally random learning mechanism.

“It is highly efficient since it is a globally ordered and locally random learning mechanism [7].”

• cites [7] Zhang, 2020, Non-iterative and Fast Deep Learning: Multilayer Extreme Learning Machines, J. Frankl. Inst. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Large weight changes cause forgetting by disrupting previously learned representations, while too much stability prevents the model from learning new tasks.

“The model requires sufficient plasticity to acquire new tasks, but large weight changes will cause forgetting by disrupting previously learned representations. Keeping the network’s weights stable prevents previously learned tasks from being forgotten, but too much stability prevents the model from learning new tasks [6].”

• cites [6] Kemker, R., McClure, M., Abitino, A., Hayes, T., and Kanan, C. (2018, January 2–7). Measuring catastrophic forgetting in neural networks. Proceedings of the AAAI, New Orleans, LA, USA. sourceNot checked yet

  Not checked in the free run.

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

Incremental learning requires the model to retain its performance on the old task after learning a new task, whereas transfer learning only uses old knowledge to learn new knowledge and focuses only on new knowledge performance.

“Compared with transfer learning, incremental learning requires the model to retain its performance on the old task after learning a new task. In contrast, transfer learning only uses the old knowledge to learn new knowledge. After the learning is completed, it only focuses on the performance of the new knowledge, and no longer considers the performance of the old knowledge.”

• AssumptionRests on an assumption: The external environment of the real world is dynamically changing, requiring an intelligent agent to have continuous learning and memorizing ability.

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

The difficulty of simultaneously meeting plasticity and stability requirements is called the stability-plasticity dilemma.

“This phenomenon is called the stability-plasticity dilemma [5].”

• cites [5] Abraham, 2005, Memory retentions-the synaptic stability versus plasticity dilemma, Trends Neurosci. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

OWM and OGD update gradients based on orthogonal constraints, which is an intuitive and effective way to maintain model stability, but they still cannot avoid the limitations of the regularization strategy itself.

“OWN and OGD update gradients based on orthogonal constraints, which is an intuitive and effective way to maintain model stability. However, they still cannot avoid the limitations of the regularization strategy itself.”

• cites [51] Zeng, G.X., Chen, Y., Cui, B., and Yu, S. (2018). Continuous Learning of Context-dependent Processing in Neural Networks. arXiv.Not checked yet

  Not checked in the free run.

• cites [52] Farajtabar, M., Azizan, N., Mott, A., and Li, A. (2020, January 3). Orthogonal Gradient Descent for Continual Learning. Proceedings of the AISTATS, Palermo, Italy.Not checked yet

  Not checked in the free run.

May combine two claims.

Existing surveys on incremental learning methods are somewhat dated and lack in-depth analysis, especially for overcoming catastrophic forgetting.

“However, the existing researches are somewhat dated and lack in-depth (thorough) analysis especially for overcoming CF.”

• cites [18] Losing, 2018, Incremental on-line learning: A review and comparison of state of the art algorithms, Neurocomputing sourceNot checked yet

  Not checked in the free run.

• cites [19] Chefrour, 2019, Incremental supervised learning: Algorithms and applications in pattern recognition, Evol. Intell. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Learn++ has the ability to learn new classes, a small number of parameters, and short training time, but suffers from data imbalance when learning new classes.

“As an algorithm proposed earlier, the ability to learn new classes is one of the main features and advantages of learn++. It also has the advantages of a small number of parameters and short training time. However, it suffers from data imbalance when learning new classes.”

• cites [24] Polikar, 2001, Learn++: An incremental learning algorithm for supervised neural networks, IEEE Trans. Syst. Man Cybern Part C sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Self-Net's parameters grow only logarithmically with the number of tasks, and it has achieved good results in storage compression.

“Since parameters grow only logarithmically with the number of tasks, Self-Net has achieved good results in storage compression.”

• cites [32] Mandivarapu, 2020, Self-Net: Lifelong Learning via Continual Self-Modeling, Front. Artif. Intell. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Compared with EWC, SI measures the importance of parameters more intuitively.

“Compared with EWC, SI measures the importance of parameters more intuitively.”

• cites [46] Zenke, 2017, Continual Learning Through Synaptic Intelligence, Proc. Mach. Learn. Res.Not checked yet

  Not checked in the free run.

May combine two claims.

EWC can effectively save storage space because it alleviates catastrophic forgetting without expanding the network or retaining old data, but it only considers the Fisher information matrix for the final stage, not all previous stages, so interval forgetting can still occur.

“EWC can effectively save storage space since it is a way to alleviate catastrophic forgetting without expanding the network and retaining old data.”

• Also stated in: § 4.2. Regularization Strategy · plain

  “However, EWC only considers the Fisher information matrix for the final stage, not all the previous stages. Therefore, there will still be the phenomenon of interval forgetting.”

• cites [15] Kirkpatrick, 2016, Overcoming catastrophic forgetting in neural networks, Proc. Natl. Acad. Sci. USA sourceNot checked yet

  Not checked in the free run.

May combine two claims.

PNN fixes parameters of networks trained by previous tasks and introduces previously learned knowledge to protect old-task performance and effectively alleviate catastrophic forgetting.

“PNN retains the structure that has been trained to protect the performance of the model on the old task and effectively alleviate catastrophic forgetting.”

• cites [27] Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., and Hadsell, R. (2016). Progressive neural networks. arXiv.Not checked yet

  Not checked in the free run.

May combine two claims.

PNN's number of parameters gradually increases as tasks increase, and designing different tasks requires manual intervention.

“However, the number of parameters will gradually increase as tasks increase, and the design of different tasks requires manual intervention.”

• cites [27] Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., and Hadsell, R. (2016). Progressive neural networks. arXiv.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The dynamic expansion network (DEN) based on correlation can make more efficient use of network capacity and save storage space.

“This dynamic expansion network (DEN) based on correlation can make more efficient use of network capacity and save storage space.”

• cites [29] Yoon, J., Yang, E., Lee, J., and Hwang, S.J. (2017). Lifelong Learning with Dynamically Expandable Networks. arXiv.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Among strategies for dealing with catastrophic forgetting, the rehearsal strategy has the longest history and works well.

“Among several strategies for dealing with CF, the rehearsal strategy has the longest history and works well.”

May combine two claims.

Compared with LwF, GD strengthens the binding force of distillation loss, which helps improve model stability but also loses some efficiency.

“Compared with LwF, GD strengthens the binding force of distillation loss, which helps to improve model stability, but also loses some efficiency.”

• cites [76] Lee, K.B., Lee, K.M., Shin, J.W., and Lee, H.L. (November, January 27). Overcoming Catastrophic Forgetting With Unlabeled Data in the Wild. Proceedings of the ICCV, Seoul, Korea.Not checked yet

  Not checked in the free run.

May combine two claims.

Pseudo-rehearsal requires a generative model that can meet incremental requirements, which currently seems more difficult than implementing incremental learning.

“However, the pseudo-rehearsal strategy requires the use of a generative model that can meet the incremental requirements, which currently seems more difficult than implementing incremental learning.”

May combine two claims.

Since catastrophic forgetting can be greatly relieved by retaining a small amount of old data, rehearsal strategy often appears as an effective auxiliary method in various incremental learning methods.

“Since CF can be greatly relieved by only retaining a small amount of old data, rehearsal strategy often appears as an effective auxiliary method in various incremental learning methods.”

• AssumptionRests on an assumption: One of the reasons for catastrophic forgetting is that incremental learning lacks corresponding supervision for previous knowledge.
• AssumptionRests on an assumption: If a model can review past knowledge when learning new knowledge, it can mitigate catastrophic forgetting.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Distillation is a macro-protection oriented regularization method that constrains the output values of the old and new models, making the new model consistent with the old model when learning new data and partly overcoming catastrophic forgetting.

“Distillation is a macro-protection oriented regularization method, which constrains the output value of the old model and new model. This can make the new model consistent with the old model when learning new data, and knowledge contained in an old model can be drawn into the new model and CF can be partly overcome.”

May combine two claims.

These challenges of rehearsal could be alleviated by optimizing data storage methods, but still cannot be completely overcome.

“These challenges could be alleviated by optimizing data storage methods [99], but it still cannot be completely overcome.”

• cites [99] Riemer, M., Klinger, R., Bouneffouf, D., and Franceschini, M. (February, January 27). Scalable Recollections for Continual Lifelong Learning. Proceedings of the AAAi, Honolulu, HI, USA.Not checked yet

  Not checked in the free run.

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

Compared with other methods, MEDIC has conducted a more comprehensive and detailed evaluation of model performance, including average task accuracy, forgetting measure and intransigence.

“Compared with other methods, MEDIC has conducted a more comprehensive and detailed evaluation of the performance of the model, including: average task accuracy, forgetting measure and intransigence.”

• cites [73] Kim, D., Bae, J., Jo, Y., and Choi, J.Y. (2019). Incremental Learning with Maximum Entropy Regularization: Rethinking Forgetting and Intransigence. arXiv.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

Incremental learning is still a hot research area and will be for a long period.

“It is concluded that incremental learning is still a hot research area and will be for a long period.”

• AssumptionRests on an assumption: Incremental learning remains a long term challenge.

May combine two claims.

Incremental learning algorithms based on the architectural strategy usually require task identity in advance during inference, which restricts robustness and flexibility and makes them difficult for more realistic environments.

“In addition, the incremental learning algorithm based on the architectural strategy usually requires the task identity to be informed in advance during the inference phase [27,29], which restricts the robustness and flexibility of the strategy and makes it difficult to tackle for a more realistic environment.”

• cites [27] Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., and Hadsell, R. (2016). Progressive neural networks. arXiv.Not checked yet

  Not checked in the free run.

• cites [29] Yoon, J., Yang, E., Lee, J., and Hwang, S.J. (2017). Lifelong Learning with Dynamically Expandable Networks. arXiv.Not checked yet

  Not checked in the free run.

May combine two claims.

Different tasks may be sensitive to the same parameters, and processing this is a big challenge.

“Furthermore, different tasks may be sensitive to the same parameters, and the processing method at this time is also a big challenge.”

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

As each strategy has limitations, researchers increasingly tend to combine various strategies to achieve incremental learning.

“As each strategy has corresponding limitations, researchers increasingly tend to combine various strategies to achieve incremental learning.”

May combine two claims.

It is difficult for current generators to generate complex pictures, so pseudo-rehearsal can only achieve results in some simple classification problems.

“Moreover, it is difficult for the current generator to generate complex pictures, which means the pseudo-rehearsal strategy can only achieve results in some simple classification problems.”

May combine two claims.

Compared with rehearsal, pseudo-rehearsal does not need to save a large number of real samples and has advantages in protecting privacy and saving memory.

“Comparing with the rehearsal strategy, the pseudo-rehearsal strategy does not need to save a large number of real samples, and has great advantages in protecting privacy and saving memory.”

May combine two claims.

If memory capacity is limited, the sample size of a single knowledge category will gradually decrease with task accumulation, and its impact on the model will gradually decrease.

“If the memory capacity is limited [17], the sample size of a single knowledge category will gradually decrease with the accumulation of tasks, and its impact on the model will gradually decrease.”

• cites [17] Rebuffi, S.A., Kolesnikov, A., Sperl, G., and Lampert, C.H. (2017, January 21–26). iCaRL: Incremental classifier and representation learning. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. sourceNot checked yet

  Not checked in the free run.

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

The KT method requires an equal number of categories in the old and new tasks.

“However, this requires an equal number of categories in the old and new tasks.”

• cites [75] Jung, H.C., Ju, J.W., Jung, M.J., and Kim, J.M. (2018, January 2–7). Less-Forgetful Learning for Domain Expansion in Deep Neural Networks. Proceedings of the AAMI, New Orleans, Louisiana, LA, USA. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Multi-head classifier requires task identity in advance during inference, which limits its versatility.

“However, the multi-head classifier requires the task identity to be informed in advance during the inference stage, which is a strict condition that limits its versatility.”

May combine two claims.

The experimental protocol of some methods is closer to instance incremental scenario, and they obtained relatively good results; this shows instance incremental scenario is less difficult than class-incremental scenario.

“The experimental protocol of some methods is closer to the instance incremental scenario, and all have obtained relatively good experimental results [15,33]. It shows that the implementation of an instance incremental scenario is less difficult than the class-incremental scenario.”

• cites [15] Kirkpatrick, 2016, Overcoming catastrophic forgetting in neural networks, Proc. Natl. Acad. Sci. USA sourceNot checked yet

  Not checked in the free run.

• cites [33] Yang, Y., Zhou, D.W., Zhan, D.C., Xiong, H., and Jiang, Y. (2019, January 4–8). Adaptive Deep Models for Incremental Learning: Considering Capacity Scalability and Sustainability. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, Anchorage, AK, USA. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The main drawback of the rehearsal strategy is that storing old data requires a lot of memory space, and there is not infinite space to store enough data.

“The main drawback of the rehearsal strategy is that storing old data requires a lot of memory space, but we cannot have infinite space to store enough data.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The thought of reducing plasticity of neural networks to improve stability in the regularization strategy is theoretically close to the long-term memory of biological brains.

“In terms of the regularization strategy, the thought of reducing the plasticity of neural networks to improve stability is theoretically close to the long-term memory of biological brains.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

With the maturity of generative adversarial networks, the pseudo-rehearsal strategy has recently received increasing attention.

“With the maturity of generative adversarial networks, the pseudo-rehearsal strategy has recently received increasing attention.”

May combine two claims.

Weight regularization does not require extra storage space and sometimes has a good effect on improving network stability and mitigating catastrophic forgetting.

“Weight regularization does not require extra storage space and sometimes has a good effect on improving network stability and mitigating catastrophic forgetting.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

In many cases, due to security and personal privacy, old data is not allowed to be retained.

“In addition, in many cases, due to considerations such as security and personal privacy, old data is not allowed to be retained.”

May combine two claims.

EWC can be used in class-incremental scenario, indicating regularization strategy is applicable to both instance and class-incremental scenarios.

“In addition, EWC, as a representative method of weight regularization, can also be used in a class-incremental scenario, which indicates that the regularization strategy is applicable to both an instance incremental scenario and class-incremental scenario.”

May combine two claims.

Compared with LwF, AFA improves the distillation loss and increases the restraint of the model, but it does not significantly improve the performance of the model.

“Compared with LwF, AFA improves the distillation loss and increases the restraint of the model, but it does not significantly improve the performance of the model.”

• cites [71] Yao, 2019, Adversarial Feature Alignment: Avoid Catastrophic Forgetting in Incremental Task Lifelong Learning, Neural Comput. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

In knowledge distillation regularization, supervision of soft labels from new data is not strong enough, and there is an imbalance between old and new classes.

“However, the issue that follows is that the supervision of soft labels obtained using new data is not strong enough, and there will be an imbalance between the old and new classes.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Since current research is almost task-based, experimental protocol is strict and cannot simulate real environment well.

“Since the current research on incremental learning is almost task-based, the experimental protocol is strict, which cannot simulate the real environment very well.”

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

Before deep learning, Robins stated that catastrophic forgetting could be mitigated through rehearsal or pseudo-rehearsal.

“Before the era of deep learning, Robins [86,87] stated that catastrophic forgetting could be mitigated through rehearsal or pseudo-rehearsal.”

• cites [86] Anthony, R. (1993, January 24–26). Catastrophic forgetting in neural networks: The role of rehearsal mechanisms. Proceedings of the New Zealand International Two-stream Conference on Artificial Neural Networks & Expert Systems IEEE, Dunedin, New Zealand.Not checked yet

  Not checked in the free run.

• cites [87] Robins, 1995, Catastrophic forgetting, rehearsal and pseudorehearsal, Connect. Sci. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Knoblauch et al. pointed out that IL algorithms can be seen as polynomial time heuristics targeted at solving an NP-HARD problem and theoretically revealed why rehearsal and pseudo-rehearsal can more effectively alleviate catastrophic forgetting.

“Recently, a study by Knoblauch et al. pointed out that IL algorithms can be seen as polynomial time heuristics targeted at solving an NP-HARD problem and theoretically revealed why a rehearsal and pseudo-rehearsal strategy can more effectively alleviate catastrophic forgetting [88].”

• cites [88] Knoblauch, J., Husain, H., and Diethe, T. (2020). Optimal Continual Learning has Perfect Memory and is NP-hard. arXiv.Not checked yet

  Not checked in the free run.

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

The universal applicability of incremental learning makes it occupy an important position in other machine learning fields, such as incremental transfer learning and incremental reinforcement learning.

“The universal applicability of incremental learning makes it occupy an important position in the research of other machine learning fields, such as incremental transfer learning [102], incremental reinforcement learning [115], and so on.”

• cites [102] Lopez-Paz, D., and Marc’Aurelio, R. (2017, January 4–9). Gradient episodic memory for continual learning. Proceedings of the NIPS, Long Beach, CA, USA.Not checked yet

  Not checked in the free run.

• cites [115] Wang, 2020, Incremental Reinforcement Learning in Continuous Spaces via Policy Relaxation and Importance Weighting, IEEE Trans. Neural Netw. Learn. Syst. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

In real environment, data flows constantly and there is no obvious boundary between data, i.e., task boundaries will not be predefined.

“In the real environment, data flows are coming constantly, and there is no obvious boundary between the data, i.e., the boundaries of the task will not be predefined.”

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

The research reveals that current application of incremental learning in assistive robots is still far from enough, and algorithms that can tackle these practical factors urgently need to be developed.

“The research reveals that the current application of incremental learning in assistive robots is still far from enough, and algorithms that can tackle these practical factors urgently need to be developed.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Lifelong SLAM emphasizes positioning failure and mismatch problems caused by scene changes, which could be addressed by incremental learning.

“Lifelong SLAM emphasizes the positioning failure and mismatch problems caused by scene changes, which could be addressed by incremental learning.”

• cites [137] Shi, X.S., Li, D.J., Zhao, P.P., Tian, Q.B., Tian, Y.X., Long, Q.W., Zhu, C.H., Song, J.W., Qiao, F., and Song, L. (August, January 31). Are We Ready for Service Robots? The OpenLORIS-Scene Datasets for Lifelong SLAM. Proceedings of the 2020 IEEE International Conference on Robotics and Automation (ICRA), Paris, France. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The proportion of attempts to combined strategy is gradually increasing, implying current research on incremental learning has fallen into a bottleneck.

“What is more, the proportion of attempts to the combined strategy is gradually increasing, which implies that the current research on incremental learning has fallen into a bottleneck.”

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

Using a combination of various existing strategies for incremental learning is usually better than using a single strategy alone.

“Researchers are increasingly inclined to use a combination of various existing strategies to study incremental learning, which is usually better than using a single strategy alone.”

Probably not a substantial claim: may not pass S2 (checkable).

The realization of flexible and stable incremental learning that can adapt to various complex environments is still far away.

“Although the current work has made good progress, the realization of flexible and stable incremental learning that can adapt to various complex environments is still far away.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Many articles have emphasized that catastrophic forgetting does not occur in the human brain, but have ignored that forgetting can also occur in the human brain.

“For incremental learning, many articles have emphasized that catastrophic forgetting does not occur in the human brain, but have ignored the fact that forgetting can also occur in the human brain.”

• AssumptionRests on an assumption: Even if long-term memory is formed by creating new synapses, the human brain will forget it after a long period of inactivity.

May combine two claims.

FearNet directly uses pre-trained ResNet embeddings as extracted features, making its ability to extract features in new data questionable.

“However, it directly uses pre-trained ResNet embeddings as extracted features that feed to FearNet, which makes the model’s ability to extract features in new data questionable.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

FearNet's dual memory system conforms to mammalian memory system and is a model design method worth exploring.

“The dual memory system designed by FearNet [92] conforms to the mammalian memory system and is a model design method worth exploring.”

• cites [92] Kemker, R., and Kanan, C. (May, January 30). FearNet: Brain-Inspired Model for Incremental Learning. Proceedings of the ICLR, Vancouver, BC, Canada.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Regularization strategies are attractive because they are efficient and require no extra storage.

“Regularization strategies are attractive since they are efficient and require no extra storage.”

May combine two claims.

Using similar experimental settings, the gap between results of three distillation-based methods is not obvious; this means soft label supervision has not been significantly enhanced, and using only new data is not reliable at present.

“Using similar experimental settings, the gap between the results of the three methods based on the distillation strategy is not too obvious [16,70,71]. This means that the supervision ability of soft labels has not been significantly enhanced, i.e., the method of using only new data is not reliable at present.”

• cites [16] Li, 2016, Learning Without Forgetting, ECCVNot checked yet

  Not checked in the free run.

• cites [70] Hou, S.H., Pan, X.Y., Chen, C.L., Wang, Z.L., and Lin, D.H. (2018, January 8–14). Lifelong Learning via Progressive Distillation and Retrospection. Proceedings of the ECCV, Munich, Germany. sourceNot checked yet

  Not checked in the free run.

• cites [71] Yao, 2019, Adversarial Feature Alignment: Avoid Catastrophic Forgetting in Incremental Task Lifelong Learning, Neural Comput. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Existing incremental learning algorithms generally suffer from poor flexibility and strict environmental requirements.

“In addition to being a bottleneck in strategy, the existing incremental learning algorithms generally suffer from poor flexibility and strict environmental requirements.”

May combine two claims.

The rehearsal strategy can make up for the lack of soft label supervision in distillation, and distillation can improve training efficiency and reduce space consumption.

“On the other hand, the rehearsal strategy can make up for the lack of soft label supervision in the distillation strategy, and the distillation strategy can improve training efficiency and reduce space consumption.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The pseudo-rehearsal strategy has more possibilities due to the development of GANs.

“The pseudo-rehearsal strategy has more possibilities due to the development of GANs.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Although iCaRL limits memory consumption to some extent, it does not meet the requirements of long-term increments.

“Although iCaRL limits memory consumption to some extent, it does not meet the requirements of long-term increments.”