Catastrophic Forgetting in Deep Learning: A Comprehensive Taxonomy — part 9/9

Cleaned-text code points [123878, 132713). Bibliography omitted here; source.md includes it.

The primary challenge encountered in the distance-based category is maintaining the consistency of the class representations when new training sessions occur. As new tasks or data are introduced, it is crucial to ensure that the representations remain reliable and informative. Overcoming this challenge is essential for preventing CF and enabling effective classification on both old and new tasks. To enhance the distance-based approaches, future research could focus on developing more robust and accurate embedding generators that can better capture the subtle differences between classes. Additionally, investigating novel unsupervised learning techniques and refining ranking loss functions can further improve the effectiveness of this category. By addressing these challenges, we can enhance the performance and stability of distance-based methods for mitigating catastrophic forgetting.

The third category we reviewed was sub-networks, which addresses the issue of CF by minimizing changes in neuron connections during the learning of new tasks. In Sub-network the CF problem arises when the value of a neuron connection, which has already learned a task t i subscript 𝑡 𝑖 t_{i} italic_t start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT , changes in favor of learning a new task t i + 1 subscript 𝑡 𝑖 1 t_{i+1} italic_t start_POSTSUBSCRIPT italic_i + 1 end_POSTSUBSCRIPT . Thus, the main objective of sub-networks approaches is to mitigate such changes. Notably, Elastic Weight Consolidation (EWC) by ? (?) and PathNet by ? (?) are pioneering works in this category, although they may not achieve the best performance results.

We can classify similar approaches to EWC as Soft Sub-networks since they allow changes in all parameters of the neural network while penalizing changes in connections that are important for previous tasks. On the other hand, works similar to PathNet fall into the category of Hard Sub-networks, where the values of neuron connections, specifically those crucial for learning certain tasks, are frozen after training to prevent catastrophic forgetting. However, it is important to consider the limitations of Hard Sub-networks, such as their restricted number of parameters, which restricts the total number of tasks that can be learned. Additionally, the usage of an oracle is necessary for PathNet to determine the parameters to be used for each sample during inference. In contrast, Soft Sub-networks do not face this limitation, but they show lower performance in scenarios where domain tasks drift. Therefore, it is crucial to explore improved techniques for selecting parameters as an oracle in such contexts, as it holds the potential for significant advancements in the sub-networks category.

The final category we reviewed was dynamic networks, which aims to address the limitations of limited resources by dynamically increasing the size of the neural network as required. The representative work in this category is Progressive Neural Networks (PNN) by ? (?), which overcome constraints by allocating more resources for each new task. For instance, PNN demonstrates the creation of a completely new neural network for each task, resulting in improved performance.

The key research contribution in this category is to determine when and how to expand the network structure effectively. Finding optimal strategies for increasing the network’s capacity is crucial. Additionally, exploring approaches to reduce the network structure could also lead to significant improvements. This is because as we gain knowledge in solving tasks and subsequently learn new incremental tasks, the previously learned tasks tend to become easier. However, a common challenge encountered here is the selection of parameters for inference. Determining the appropriate parameters set for a given sample during inference is critical for achieving optimal performance. Hence, it is recommended to investigate techniques that can serve as an oracle for parameter selection. Advancements in developing effective oracle techniques would benefit both the dynamic networks and sub-networks categories.

Lastly, hybrid models offer a promising approach to mitigating CF. Rehearsal techniques can be effectively combined with methods from other categories to address this challenge. However, there is a lack of comparative studies on the performance of these combined approaches. The combination of sub-network and dynamic network methods may not seem intuitive at first, as the former focuses on identifying task-specific sub-structures within the model, while the latter expands the model’s capacity to accommodate new knowledge. Nevertheless, integrating elements from different categories can help reduce resource limitations encountered by sub-networks. In a study by (?), a mixture of techniques in the form of ensembles, such as EWC, IMM, and ICaRL, was explored. This approach aims to leverage the strengths of each model to enhance resistance against CF. However, it is worth noting that ensembles can come with a high computational cost.

To date, our survey has examined a comprehensive selection of over 230 studies, primarily focusing on works published after 2013 when Deep Learning gained significant attention. We aimed to capture the most recent advancements in the field. Figure 16 presents the distribution of works across the proposed categories over the years. Notably, during the initial period from 2013 to 2015, the Sub-networks category dominated the landscape as the primary focus for mitigating CF. However, in subsequent years, we observed a shift in research interest, with increasing contributions from other categories, resulting in a more balanced distribution. This diverse distribution implies that there is no clear-cut path or prevailing approach to achieving continuous learning while effectively addressing the challenge of CF.

We have observed that in the last two years, a majority of the works have adopted some variant of the Mini-rehearsal technique. It is worth noting that several studies have utilized techniques similar to a coreset to address CF, such as Incremental Semantic Segmentation (?, ?). However, during our review, we discovered that not all papers proposed novel techniques specifically targeting CF. Instead, many of them applied existing techniques to various applications, making it impractical to cite each individual work. To overcome this limitation, we recommend referring to a curated repository 1 1 1 https://github.com/xialeiliu/Awesome-Incremental-Learning that provides an extensive and regularly updated list of papers on Incremental Learning, where the majority of them attempt to address CF. In this review, our focus was on works that have made significant contributions to each evolving category.

More recently, in the domains of Natural Language Processing and Computer Vision, the utilization of prompts combined with few-shot learning in Transformers (ViT) has shown promise in addressing incremental learning without retraining the model, thereby mitigating the risk of forgetting (?, ?, ?, ?, ?). However, it is important to acknowledge that this approach is not a definitive solution to CF, as its effectiveness is confined to specific domains. Furthermore, the current computational demands of most transformer models make them unsuitable for regular updates of their parameters to solve new tasks on a regular basis. Furthermore, it should be noted that most of these proposals present good results only in the context of Task Incremental Learning, making their practical implementation unfeasible in more realistic scenarios, such as Online or Unbound Incremental Learning.

Although catastrophic forgetting (CF) is a well-established problem, there remains considerable divergence in testing protocols. Therefore, it is crucial for the field to adopt a standardized test platform that enables fair comparisons among new models and existing ones (?). As a final consideration, it is important to recognize that the choice of the best model, technique, or framework will depend on the specific context and problem being addressed. Furthermore, it is likely that solving new complex problems will require the integration of multiple categories rather than relying solely on a single approach.

Acknowledgments

The present work is the result of the Research and Development (R&D) project 001/2020, signed with Federal University of Amazonas and FAEPI, Brazil, which has funding from Samsung, using resources from the Informatics Law for the Western Amazon (Federal Law nº 8.387/1991), and its disclosure is in accordance with article 39 of Decree No. 10.521/2020. And, is a result of the R&D project (Tracking Platform), carried out by Sidia Instituto de Ciência e Tecnologia, in partnership with Samsung Eletrônica da Amazônia Ltda.