CoChat
Sign Up
Cited by 259Domain Adaptation and Few-Shot LearningAdvanced Neural Network ApplicationsData Stream Mining Techniques

Tree-CNN: A hierarchical Deep Convolutional Neural Network for incremental learning

Open access
Roy, Deboleena, Panda, Priyadarshini, Roy, Kaushik·2020·Neural Networks
DOI 10.1016/j.neunet.2019.09.010
Open this paper in CoChat — free account

Abstract

Over the past decade, Deep Convolutional Neural Networks (DCNNs) have shown remarkable performance in most computer vision tasks. These tasks traditionally use a fixed dataset, and the model, once trained, is deployed as is. Adding new information to such a model presents a challenge due to complex training issues, such as "catastrophic forgetting", and sensitivity to hyper-parameter tuning. However, in this modern world, data is constantly evolving, and our deep learning models are required to adapt to these changes. In this paper, we propose an adaptive hierarchical network structure composed of DCNNs that can grow and learn as new data becomes available. The network grows in a tree-like fashion to accommodate new classes of data, while preserving the ability to distinguish the previously trained classes. The network organizes the incrementally available data into feature-driven super-classes and improves upon existing hierarchical CNN models by adding the capability of self-growth. The proposed hierarchical model, when compared against fine-tuning a deep network, achieves significant reduction of training effort, while maintaining competitive accuracy on CIFAR-10 and CIFAR-100.

AI snapshot

Written from the paper’s full text.

TL;DR

Tree-CNN is a hierarchical DCNN that grows tree-like as new classes arrive, assigning new classes to existing, merged, or new child nodes based on softmax likelihood and retraining only affected nodes. On CIFAR-10 and CIFAR-100 it reduces training effort versus fine-tuning a baseline network while maintaining competitive accuracy.

Key findings

  • On CIFAR-10, after adding 4 new classes, Tree-CNN reached 86.24% test accuracy with normalized training effort 0.60, versus baseline B:V at 90.51% accuracy and effort 1.00; Tree-CNN used 40% less effort than B:V and 30% less than B:II, while being 3.76% below B:V.
  • On CIFAR-100 over 9 incremental stages, Tree-CNN-5/10/20 achieved final test accuracies of 61.57%, 60.46%, and 59.99% (average 69.85%, 69.53%, 68.49%), compared with B:V at 63.05% final / 72.23% average, iCaRL at 49.11% / 64.10%, and LwF at 25.00% / 44.49%. The paper reports Tree-CNNs yield 10% higher accuracy than iCaRL and over 50% higher than LwF.
  • Tree-CNNs required lower training effort than fine-tuning cases B:II–B:V; B:I had the lowest effort but the worst accuracy degradation. Tree-CNN-5 grew to 3.4× the size of Network B, Tree-CNN-10 to 2.2×, and Tree-CNN-20 to 1.8×.
  • The initial CIFAR-10 Tree-CNN achieved 98.73% root-node testing accuracy, 86% for the “Animals” branch, 94.43% for “Vehicles,” and 89.10% overall.

Methods

  • Tree-CNN is composed of DCNNs in a tree: the root classifies inputs into child nodes, branch nodes perform finer classification, and each leaf node is uniquely associated with one class. Incremental learning samples 10% of new-class training images, computes a softmax likelihood matrix, orders classes, then adds each new class to an existing child, merges child nodes, or adds a new child; affected nodes are retrained with old and new data.
  • Experiments used CIFAR-10 (6 initial classes, then 4 new classes) and CIFAR-100 (100 classes, 500 training and 100 testing images per class, randomly divided into 10 groups of 10 and introduced over 9 stages). CIFAR-100 used max depth 2, thresholds 0.1, and maxChildren 5, 10, or 20. Baseline Network B was a VGG-like 11-layer network fine-tuned at five depths (B:I–B:V).
  • Metrics were testing accuracy and training effort, defined as the sum over nets of total weights × total training samples. Training used MatConvNet, mini-batch SGD with momentum 0.9, dropout, batch normalization, weight decay 0.001, 300 epochs, and learning rate 0.1 for 200 epochs then reduced by 10 every 50 epochs.

Limitations

  • The paper states Tree-CNN continues to grow in size over time, and the implications for memory requirements need investigation; it also notes there is no standardized benchmark protocol for incremental learning, so it used a protocol similar to iCaRL.
  • The paper states branch nodes have a higher chance of overfitting as dataset size per node shrinks, so more dropout layers were added; it also notes that if the root node misclassifies an input, the final classification is incorrect regardless of the branch node’s output, and old data must be shown during retraining to avoid catastrophic forgetting.
  • Evident from the text: evaluation is limited to CIFAR-10 and CIFAR-100; final CIFAR-100 accuracy for Tree-CNN-5 (61.57%) remains below B:V (63.05%) and below HD-CNN (67.38%) and Hertel et al. (67.68%) as listed in Table 7.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.neunet.2019.09.010

Cites

38 of 38 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 13 cited sources checked · 0 supported · 0 flagged

  • Findingreasoning
    § 5.2. Sequentially adding new classes (CIFAR-100) · p. 17

    The size of Network B remains relatively unchanged as new classes are added, because only additional output nodes are added, translating to a small fraction of new weights in the final layer.

    “The size of Network B remains relatively unchanged, as only additional output nodes are added, which translates to a small fraction of new weights in the final layer.”
    Quote in the paperSubstantial claim
  • Findingown data
    § 5.2. Sequentially adding new classes (CIFAR-100) · p. 18

    B:I suffers the worst accuracy degradation over the 9 learning stages.

    “However, it suffers the worst accuracy degradation over the 9 learning stages (Fig. 6b).”
    Quote in the paperSubstantial claim
  • Findingown data
    § Abstract · p. 1

    The proposed hierarchical model maintains competitive accuracy on CIFAR-10 and CIFAR-100.

    “The proposed hierarchical model, when compared against fine-tuning a deep network, achieves significant reduction of training effort, while maintaining competitive accuracy on CIFAR-10 and CIFAR-100.”
    Quote in the paperSubstantial claim
  • Findingown data
    § 5.2. Sequentially adding new classes (CIFAR-100) · p. 21

    At the end of the nine incremental learning stages, certain similar objects were grouped together for Tree-CNN-10.

    “At the end of the nine incremental learning stages, certain similar objects grouped together is shown in Fig. 8 for Tree-CNN-10 .”
    Quote in the paperNot a substantial claim
  • Findingown data
    § 5.2. Sequentially adding new classes (CIFAR-100) · p. 17

    As new classes are added, the Tree-CNNs grow in size by adding more branches.

    “We observe that as new classes are added, the Tree-CNNs grow in size by adding more branches (Fig. 5a).”
    Quote in the paperSubstantial claim
  • May combine two claims. 23
  • Classified as probably not a substantial claim. 12

  • May combine two claims.

    Findingown data
    § 5.2. Sequentially adding new classes (CIFAR-100) · p. 17

    Tree-CNN-5 grows to almost 3.4 times the size of Network B, while Tree-CNN-10 and Tree-CNN-20 reach 2.2 and 1.8 times the baseline size, respectively.

    “Tree-CNN-5 almost grows 3 : 4 the size of Network B , while Tree-CNN-10 and Tree-CNN-20 reach 2 : 2 and 1 : 8 the baseline size, respectively.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.1.2. The Network Initialization · p. 11

    The branch nodes 'Animals' and 'Vehicles' achieve 86% and 94.43% testing accuracy respectively on CIFAR-10.

    “The root node achieves a testing accuracy of 98 : 73% , while the branch nodes, 'Animals' and 'Vehicles', achieve 86% and 94 : 43% testing accuracy respectively.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 5.2. Sequentially adding new classes (CIFAR-100) · p. 21

    There were some groups that had objects sharing semantic similarity, and there were also odd groups, such as Node 13.

    “While there were some groups that had object sharing semantic similarity as well, there were odd groups as well, such as Node 13 as shown in Fig. 8.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 5.2. Sequentially adding new classes (CIFAR-100) · p. 19

    The three Tree-CNNs perform almost at par with B:IV, and outperform all other variants of network B.

    “The three Tree-CNNs perform almost at par with B:IV , and outperform all other variants of network B .”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 5.1. Adding multiple new classes (CIFAR-10) · p. 17

    Tree-CNN has the second lowest normalized training effort for CIFAR-10, 40% less than B:V and 30% less than B:II.

    “The Tree-CNN has the second lowest normalized training effort, 40% less than B:V , and 30% less than B:II .”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 5.1. Adding multiple new classes (CIFAR-10) · p. 17

    Tree-CNN had comparable accuracy to B:II and B:III for CIFAR-10, while being less than B:V by a margin of 3.76%.

    “At the same time, Tree-CNN had comparable accuracy to B:II and B:III , while just being less than the ideal case B:V by a margin of 3 : 76% .”
    Quote in the paperSubstantial claim

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

    Premisecitation
    § 1. Introduction · p. 2

    Deep Convolutional Neural Networks have emerged as the leading architecture for large scale image classification.

    “In recent years Deep Convolutional Neural Networks (DCNNs) have emerged as the leading architecture for large scale image classification [1].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [1] Aljundi, 2016, Expert gate: Lifelong learning with a network of experts, CoRR

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning
    § 1. Introduction · p. 2

    In Tree-CNN, initial nodes assign the input into coarse super-classes, and deeper nodes perform finer classification.

    “The initial nodes of the Tree-CNN assign the input into coarse super-classes, and as we approach the leaves of the network, finer classification is done.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given
    § Algorithm 1 Tree-CNN: At Inference · p. 8

    The Tree-CNN algorithm can potentially be extended to deeper Tree-CNN models.

    “While this example is for a two level Tree-CNN, the algorithm can potentially be extended to deeper Tree-CNN models.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 2. Related Work · p. 3

    In prior hierarchical CNN works, the dataset is fixed from the beginning and prior knowledge of all classes and their properties is used to build the hierarchical model.

    “However, in these works, the dataset is fixed from the beginning, and prior knowledge of all the classes and their properties is used to build a hierarchical model.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [22] Sarwar, S. S., Ankit, A., & Roy, K. (2017). Incremental learning in deep convolutional neural networks using partial network sharing, arXiv preprint arXiv:1712.02719.

      Not checked in the free run.

    • cites [23] Sarwar, 2017, Gabor Filter assisted energy efficient fast learning convolutional neural networks

      Not checked in the free run.

    • cites [24] Shmelkov, 2017, Incremental learning of object detectors without catastrophic forgetting

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Interpretationnone given
    § Algorithm 1 Tree-CNN: At Inference · p. 8

    The decision on how to grow the Tree-CNN is semi-supervised: the algorithm decides how to grow the tree given user constraints.

    “The decision on how to grow the tree is semi-supervised: the algorithm itself decides how to grow the tree, given the constraints by the user.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given
    § 3.1. Network Architecture · p. 4

    In Tree-CNN, each leaf node is uniquely associated to a class and no two leaf nodes have the same class.

    “Each leaf node is uniquely associated to a class and no two leaf nodes have the same class.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § 2. Related Work · p. 4

    By localizing the change to a small section of the whole network, Tree-CNN reduces training effort and complexity.

    “Images belonging to the older dataset are required during retraining, but by localizing the change to a small section of the whole network, our method tries to reduce the training effort and complexity.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation
    § 2. Related Work · p. 3

    Initial layers of a CNN learn very generic features.

    “It has been observed that initial layers of a CNN learn very generic features [17] that has been exploited for transfer learning [18]' [19].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [17] Panda, 2017, Semantic driven hierarchical learning for energy-efficient image classification

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data
    § 5.2. Sequentially adding new classes (CIFAR-100) · p. 18

    Only retraining the final linear classifier (fully connected layer) is not sufficient; the feature extractors (convolutional blocks) also need to be trained on the new data.

    “This shows that only retraining the final linear classifier, i.e. the fully connected layer, is not sufficient. We need to train the feature extractors, i.e. convolutional blocks, as well on the new data.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning
    § 6. Discussion · p. 22

    Tree-CNN can learn with lower training effort than fine-tuning a deep network, while preserving much of the accuracy.

    “Thus, it can learn with lower training effort than fine-tuning a deep network, while preserving much of the accuracy.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given
    § 4.3.1. Baseline Network · p. 15

    The baseline Network B has a complexity level similar to two stage Tree-CNN.

    “To compare against the proposed Tree-CNN , we defined a baseline network (Network B ) with a complexity level similar to two stage Tree-CNN .”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning
    § 6. Discussion · p. 22

    Such a method of training and classification can be used to hierarchically classify large datasets.

    “Such a method of training and classification can be used to hierarchically classify large datasets.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning
    § Algorithm 1 Tree-CNN: At Inference · p. 8

    Once new classes are allotted locations, supervised gradient descent training is performed only on modified/new nodes, so only affected portions of the network require retraining/fine-tuning.

    “Once the new classes are allotted locations in the tree, supervised gradient descent based training is performed on the modified/new nodes. This saves us from modifying the whole network, and only affected portions of the network require retraining/fine-tuning.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given
    § 6. Discussion · p. 22

    The final leaf nodes and the distance between them can be used as a measure of how similar any two images are.

    “The final leaf nodes, and the distance between them can also be used as a measure of how similar any two images are.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § Abstract · p. 1

    The network organizes incrementally available data into feature-driven superclasses and improves upon existing hierarchical CNN models by adding the capability of self-growth.

    “The network organizes the incrementally available data into feature-driven superclasses and improves upon existing hierarchical CNN models by adding the capability of self-growth.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § 5.2. Sequentially adding new classes (CIFAR-100) · p. 19

    In iCaRL, a ResNet-32 is retrained with new data and only 2000 samples of old data (exemplars) at every stage; it recovers good performance compared to LwF but still falls short of state-of-the-art by 18%.

    “In 'iCarl' [16], a ResNet-32 is retrained with new data and only 2000 samples of old data (called 'exemplars') at every stage. It is able to recover a good amount performance, compared to 'LwF' but still falls short of state-of-the-art by 18%.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [16] Panda, 2017, FALCON: Feature driven selective classification for energy-efficient image recognition, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning
    § 6. Discussion · p. 22

    The hierarchical node-based learning model of Tree-CNN attempts to confine the change in the model to a few nodes only, thereby limiting computation costs of retraining while using all previous data.

    “The hierarchical node-based learning model of the Tree-CNN attempts to confine the change in the model to a few nodes only. And, in this way, it limits the computation costs of retraining, while using all the previous data.”
    Quote in the paperSubstantial claim

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

    Premisecitation
    § 2. Related Work · p. 3

    Hierarchical CNN models have been shown to perform at par or even better than standard DCNNs.

    “Such hierarchical CNN models have been shown to perform at par or even better than standard DCNNs [22].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] Sarwar, S. S., Ankit, A., & Roy, K. (2017). Incremental learning in deep convolutional neural networks using partial network sharing, arXiv preprint arXiv:1712.02719.

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning
    § Handling input labels inside the Tree-CNN · p. 10

    Every class is finally associated with a unique leaf node, hence leaf nodes do not require a look up table.

    “Every class is finally associated with a unique leaf node, hence leaf nodes do not require a look up table.”
    Quote in the paperNot a substantial claim

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

    Study design and conductreasoning
    § 4.3.3. Evaluation Metrics · p. 15

    Training Effort attempts to capture the number of weight updates that happen per training epoch.

    “Training Effort attempts to capture the number of weight updates that happen per training epoch.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premiseown data
    § 5.2. Sequentially adding new classes (CIFAR-100) · p. 19

    For LwF, a ResNet-32 is retrained exclusively on new data at every stage, and hence it suffers the most accuracy degradation.

    “For 'LwF', a ResNet-32 [32] is retrained exclusively on new data at every stage. Hence it suffers the most accuracy degradation.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [32] Yosinski, 2014, How transferable are features in deep neural networks?

      Not checked in the free run.

  • May combine two claims.

    Interpretationnone given
    § Abstract · p. 1

    The network grows in a tree-like fashion to accommodate new classes of data while preserving the ability to distinguish previously trained classes.

    “The network grows in a tree-like fashion to accommodate new classes of data, while preserving the ability to distinguish the previously trained classes.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given
    § 2. Related Work · p. 4

    Tree-CNN starts out as a single root node and generates new hierarchies to accommodate new classes.

    “In our work, Tree-CNN starts out as a single root node and generates new hierarchies to accommodate the new classes.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning
    § Algorithm 1 Tree-CNN: At Inference · p. 8

    At every incremental learning stage, the root node is trained on all available data because it needs to classify all objects into the new branches.

    “At every incremental learning stage, the root node is trained on all the available data as it needs to learn to classify all the objects into the new branches.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationcitation
    § 6. Discussion · p. 22

    The Tree-CNN generates hierarchical grouping of initially unrelated classes, thereby generating a label relation graph out of these classes.

    “The Tree-CNN generates hierarchical grouping of initially unrelated classes, thereby generating a label relation graph out of these classes[33].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [33]

      Not checked in the free run.

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

    Study design and conductreasoning
    § 4.3.3. Evaluation Metrics · p. 15

    As batch size and number of training epochs is kept the same, the product of the number of weights and the number of training samples used can provide a measure of the computation cost of a learning stage.

    “As batch size and number of training epochs is kept the same, the product of the number of weights and the number of training samples used can provide us with the measure of the computation cost of a learning stage.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § 3.1. Network Architecture · p. 4

    Tree-CNN is composed of multiple nodes connected in a tree-like manner, where each non-leaf node has a DCNN trained to classify the input to the node into one of its children.

    “Inspired from hierarchical classifiers, our proposed model, Tree-CNN is composed of multiple nodes connected in a tree-like manner. Each node (except leaf nodes) has a DCNN which is trained to classify the input to the node into one of it's children.”
    Quote in the paperSubstantial claim

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

    Premisecitation
    § 2. Related Work · p. 3

    Common features shared between images have been used previously to build hierarchical classifiers, either grouped semantically or feature-driven.

    “Common features, that are shared between images, have been used previously to build hierarchical classifiers. These features can be grouped semantically, such as in [20], or be feature-driven, such as 'FALCON' [21].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [20] Russakovsky, 2015, Imagenet large scale visual recognition challenge, International Journal of Computer Vision source

      Not checked in the free run.

    • cites [21] Rusu, A. A., Rabinowitz, N. C., Desjardins, G., Soyer, H., Kirkpatrick, J., & Kavukcuoglu, K., et al. (2016). Progressive neural networks, arXiv preprint arXiv:1606.04671.

      Not checked in the free run.

34 of 34

  • The paper relies on this without support.

    In the real world, data is gathered incrementally over time and all information is not available at once.

    “However, in real world, we hardly have all the information at once, and data is, instead, gathered incrementally over time.”

  • The authors state this as an assumption.

    The starting model is already trained to recognize a certain number of objects.

    “We start with the assumption that we have a model that is already trained to recognize a certain number of objects.”

  • The authors state this as an assumption.

    The system has access to all the data that has been introduced so far during fine-tuning.

    “It is assumed that the system has access to all the data that has been introduced so far.”

  • The paper relies on this without support.

    There is an absence of a standardized benchmark protocol for incremental learning.

    “There is an absence of standardized benchmark protocol for incremental learning, which led us to use a benchmarking protocol similar to one used in iCaRL [16].”

  • The paper relies on this without support.

    The dataset available to the user has unique labels assigned to each of its object classes.

    “The dataset available to the user will have unique labels assigned to each of it's object classes.”

  • The paper relies on this without support.

    Adding new image classes to a network should be easier than retraining the whole network again with all the classes.

    “The motivation of this work stems from the idea that subsequent addition of new image classes to a network should be easier than retraining the whole network again with all the classes.”

  • The paper relies on this without support.

    Catastrophic forgetting is an inherent problem in deep neural networks.

    “This is due to the inherent problem of 'catastrophic forgetting' in deep neural networks.”
7 of 7

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?

  • 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.”

    From An Appraisal of Incremental Learning Methods (2020) · cites this paper as [28]

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.

  • Tree-CNNs do not outperform every Network B variant

    Minor

    The claim includes B:V, but Table C.9 reports higher accuracy for B:V than for all three Tree-CNNs at 100 classes.

1 of 1

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Roy, D., Panda, P., & Roy, K.
Affiliations
Purdue University West Lafayette (US)
Corresponding author
Roy, Deboleena
ORCID
Venue
Neural Networks
Publisher
Elsevier BV
Published
2020-01
Volume
121
Pages
148-160
ISSN
0893-6080 (ISSN-L)
DOI
10.1016/j.neunet.2019.09.010
OpenAlex
W2804177786
Semantic Scholar
e811c76173d4b0e478d3ab5965358a6377dd48ce
Cited by
257
References
37
Citations by year
2026: 12 · 2025: 29 · 2024: 39 · 2023: 44 · 2022: 41 · 2021: 53 · 2020: 30 · 2019: 7 · 2018: 2
FWCI
15.11
Type
journal article
Language
en
Open access
Licence
Locations
  • Neural Networks (acceptedVersion)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
Semiconductor Research Corporation
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 › Artificial Intelligence › Data Stream Mining Techniques
Keywords
incremental learning, catastrophic forgetting, hierarchical convolutional neural network, class-incremental learning, adaptive network structure
MeSH terms
Deep Learning, Animals, Humans, Pattern Recognition, Automated, Photic Stimulation, Neural Networks, Computer

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 2026-10-08.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-08.

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