2.1. Knowledge Transfer and Adaptation

The real-world dynamic environment consists of multiple problems or tasks with varying situations and conditions. The system should be capable of learning knowledge, transferring it, and reusing it to perform new tasks. In addition, the system should be able to adapt to any new environment.

2.2. Control Catastrophic Forgetting

Currently, most AI systems suffer from catastrophic forgetting. A CL-based system should be capable of learning new knowledge while also retaining previously learned information. The system must be stable enough not to forget this information while being plastic enough to learn new tasks. This is known as the stability-plasticity dilemma.

2.3. Task Identification

In a real-world scenario, the system should be capable of performing multiple tasks, which requires it to identify new tasks. In addition, the system should be able to handle online learning and continuous data in the context of streaming [30].

2.4. Learning Task Similarity

The system should be able to learn the similarities among different tasks. This ability will help enhance knowledge transfer, both for forward transfer (FWT) and backward transfer (BWT). In addition, the capability to divide complex tasks into simpler tasks can also facilitate the reuse of knowledge among subtasks [31].

2.5. Noise Tolerance

Generally, AI systems are trained on controlled, well-organized, and clean datasets. In contrast, the real-world environment may contain noise, disorganized, and uncontrolled data. The system should be capable of tackling such noise, dynamic environments, and varying conditions. Various studies propose robust models; however, these models have not been studied in relation to CL.

2.6. Resource Efficiency

To mimic the CL functions of the brain, the system should manage constrained resources. New resources for every new task are undesirable. Adaptability and robustness are also key characteristics for any system to achieve CL. In real-world scenarios, the system may encounter noise, out-of-distribution data inputs, and unseen, dynamic environments. As shown in Table 1, CL is achieved in ML systems by implementing methods such as neuromodulation [32], multisensory integration [33], and episodic replay [34].

3.1. Neurogenesis

Neurogenesis is the process of generating new neurons in the central nervous system [37]. It occurs in the dentate gyrus of the hippocampal formation and in the subventricular zone of the lateral ventricles (LVs) [38, 39]. This process is expedited when the brain undergoes a variety of experiences. Additionally, changes in the size and configuration of the brain have been observed [15]. Structural connections remain almost the same throughout life [15]. During the study of the mouse brain, it was observed how newly created neuroblasts travel from the subventricular zone of the LV through the rostral migratory stream (RMS) into the olfactory bulb (OB), where mature interneuron populations are produced [40]. Although this process remains active throughout life, it occurs more actively in the early years. Some studies related to neuroimaging suggest that the infant brain is more plastic and continues to grow and develop. Biological neurons in the brain utilize constructive algorithms in the form of plasticity, involving the enlargement and pruning of synapses during neurogenesis. Several studies suggest that synaptic plasticity and neurogenesis are important mechanisms to alleviate catastrophic forgetting in the biological brain [41].

In the context of ML, the explicit inspiration from neurogenesis is achieved by various architecture-based methods, where new neurons, blocks, or branches are added to the existing network to learn new tasks [42]. In addition, the implicit analogy is that both brains and ML systems face trade-offs between stability (retaining old knowledge) and plasticity (integrating new information), leading to similar solutions like dynamic capacity expansion.

Biological neurogenesis is much more dynamic and adaptable than what current algorithms can replicate. In the brain, it is carefully controlled by factors such as environmental enrichment, stress, and chemicals like serotonin and BDNF [43].

3.2. Memory System

In the biological brain, replay is defined as the repetition of activities in some format during sleep that have already occurred while awake. This activity occurs in the hippocampus alone or collectively in the hippocampus and neocortex [44]. Similarly, studies have observed the presence of multiple memories in the brain, such as hippocampal–cortical memory [45]. It has been observed that the hippocampus is used for fast learning, whereas slow learning occurs in the cortex. This explains how the brain forms declarative memories. However, studies suggest that there may be other memory models in addition to these. The CLS theory presents the idea of two memory blocks in the brain [46]: the first is the hippocampus for fast learning and the second is the neocortex for slow and detailed learning [44].

According to the modified description of CLS, the hippocampal model continuously trains the neocortical model by replaying nonoverlapping representations of previously learned knowledge [44, 47]. Neuromodulators like acetylcholine help switch between encoding and retrieval modes. Oscillatory coupling, such as theta–gamma phase coordination, plays a key role during memory replay. Additionally, behavioral context, such as stress hormones, enhances the recall of important or emotionally significant memories. Several studies, such as [8], have proposed methods to control memory size while learning multiple tasks.

In artificial neural networks, the explicit inspiration is implemented in the form of dual-network architectures (e.g. [48]), generative replay [49] which replicate hippocampal–neocortical interactions, and replay of samples [44]. In addition, as an implicit analogy, the convergent principle is adopted through replay-based methods which address forgetting by “rehearsing” past data, mirroring the brain’s need to counteract interference with limited resources [50]. By taking inspiration from the time-varying plasticity of biological synapses, various studies have tried to replicate metaplasticity by combining memory, architecture, and regularization-based methods [24, 49, 51].

In the following two paragraphs, we will discuss two important topics related to the memory system, which are active forgetting and memory reconsolidation in brain. Active forgetting in the brain is a controlled practice in the brain. During sleep, the brain loses less important neuron connections to free memory for new learning. This process is called synaptic downscaling [52]. In addition, less important memory is identified to be eliminated through dopamine mapping [53]. Moreover, new neuron cell addition replaces the old memory cells to incorporate new learning [54].

Memory reconsolidation in the brain is the process of recalling a previously stored memory which are considered to be temporarily unstable or malleable. This process depends on the synaptic plasticity mechanism such that the memory retrieval temporarily weakens it, allowing to update for new learning [55, 56].

3.3. Neuromodulation

Studies suggest the modular nature of the human brain. It is observed that neuromodulation plays an important role in learning and adoption in different environments and contexts, knowledge remembering, augmentation and refinement [61], robustness to noise, and adjustments even during uncertainties [97]. Sparsity is closely associated with the neuromodulation. Sparse coding happens generally in the brain [98]. It is reported that the hippocampal dentate gyrus utilizes very sparse depictions for pattern discernment and knowledge retention [99]. Sparsity also offers robustness, memory capability, storing efficacy, and efficiency [100]. In this context, the explicit inspiration is the gating mechanisms in ML (e.g., [62]) which mimic neuromodulatory attention.

As shown in Table 1, brain functions such as neuromodulation, neurogenesis, and context-dependent learning are closely related to knowledge transfer. Neuromodulation is extensively inspired by various studies, especially by using architecture-based models [62].

3.4. Metaplasticity

Metaplasticity is related to the updating of synapses which depends upon the current state, history, and neural activities. It also indicates the direction, duration, and degree of synaptic plasticity in the future [16, 101]. Studies suggest that metaplasticity phenomena play various roles in learning by the biological brain such as controlling the threshold and activation for plasticity, synchronization, and consolidation of the learned knowledge [102–104]. Synaptic plasticity is the mechanism by which the brain stores the information. The capability of different synapses can be updated through neural activity [105]. Subcortical acetylcholine happens in the substantia innominata, and in the medial septum (M), dopamine happens in the ventral tegmental area and the substantia nigra compacta, noradrenaline happens in the locus coeruleus, and serotonin happens in the dorsal and medial raphe nuclei [106]. Learning new knowledge can affect the previous learning which may cause forgetting. The availability of limited resources may cause rapid forgetting. This may also occur in biological synaptic weights. Metaplasticity rules that govern the learning are the source of inspiration for the design of ANNs to cater to forgetting and solving the stability–plasticity phenomena with available resources. Based on explicit inspiration, regularization-based methods penalize changes to important weights, akin to synaptic tagging [107].

In biological metaplasticity, synaptic changes are shaped by dynamic factors like neuromodulator activity patterns such as dopamine bursts aligning with theta brain waves, and behavioral states such as sleep, which can reset plasticity through specific processes [108]. In contrast, artificial methods like elastic weight consolidation (EWC) simplify this complexity by stabilizing model weights using basic importance measures [109]. Unlike these methods, biological synapses adapt flexibly, incorporating multiple influences like sensory inputs, energy constraints, and genetic regulation to fine-tune plasticity.

3.5. Multisensor Input Learning

Biological organisms take input from multiple sensors such as vision, tactile, and auditory signals. These input signals may be noisy, distributed, nonlinear, and out of sync [110]. These signals from multiple sensors are integrated by the superior colliculus as a final input to enable final coordinated motor functions among different organisms [110]. The initiation of movements through a common motor map offers efficiency in handling these signals and responses [111]. The knowledge of the process involved in handling, filtering, and processing multisensory input from different organisms provides the lead for the task-agnostic method which is a desired property to achieve CL capability in AI systems [111].

The OB is the cortical area that receives sensory signals from the other parts of the brain and the nose. Neurons that connect directly to brain regions involved in memory, context, and emotion are mainly driven by internal states, behavioral expectations, and the context of learned odors [112]. Low-level sensitized tissues and response at the subcortical level facilitate the interface with the environment. The high-level cortical brain executes only necessary plans and selects and tunes them. This method offers better resource utilization and better learning [113]. These inputs facilitate the fast learning of tasks, associating rewards with stimuli, coordinating appropriate motor actions, and helping modulate responses [114]. As an implicit analogy, task-agnostic ML methods [86] resemble biological multisensory processing but often lack modularity seen in cortical hierarchies.

3.6. Context-Dependent Learning

Studies suggest that the context plays a pivotal role in information understanding, processing, and making inferences and strategies in the biological brain by offering in the context information [64]. Context information plays an important role in the formation of responses in the olfactory system. Similarly, the neurons related to emotion, memory, context, and information also consider context information including state and behavior in decision-making. This additive information related to the context offers a complete picture and flexibility in learning, decisions, and actions. Context modulation and gating are also involved in selective attention [115]. For example, gain modulation helps to encode target trajectories in insect vision. Studies suggest the presence of some biases that are present since birth. These inductive biases play a role in complete learning from the environment [116]. In the brain, context-dependent learning works by neuromodulatory states such that arousal controlled by norepinephrine adjusts the importance of sensory inputs dynamically, cross-regional oscillatory coupling which regulates theta–gamma brainwave interactions to link related information, and developmental metaplasticity such that the brain fine-tunes its sensitivity to context using genetic and molecular changes [113]. In comparison, ML methods like task-specific gating can handle context changes in a fixed way but lack the brain’s ability to continuously adapt and scale context sensitivity. Context-dependent learning is achieved in ML through meta-learning and task-specific gating [117]. Different studies utilize network pruning method to control the growing dynamic network [118].

4.1. Architecture-Based Methods

In this approach, architectural properties and parameters of the model are updated in response to the new task to accommodate the new information as shown in Figure 4(c). These models are inspired by modularity. Modularity is a capability of biological learning systems that defines the functional specialty of the brain subsystems. Although the working principles of neurogenesis and structural plasticity are still topics of active research, the available findings suggest that modular neural networks with plasticity offer an effective mechanism for knowledge retention and reuse even in a dynamic environment. Architecture-based methods can further be divided into dynamic networks and static networks. In both cases, the existing model is preserved and reused while adding new learning capacity. These approaches are envisaged as the knowledge retention and reuse mechanism, which is not possible to acquire by keeping the capacity of the model permanent, especially in the case of multiple dynamic tasks. Table 1 enlists the recent CL-based methods that adopt architecture-based methods for learning. It also highlights the type of network and neural network implementation methodology.

4.2. Regularization-Based Methods

Different studies suggest that synaptic plasticity takes care of the synchronization between previously learned and new knowledge [13]. Inspired by this mechanism of the brain, regularization-based methods are devised [13]. This methodology is different from architecture-based and replay-based methods in such a way that it does not generally require increasing memory or networks, rather it utilizes a regularization factor in the loss term during the training of new tasks [12] as shown in Figure 4(a).

Table 3 enlists the recent CL-based methods that adopt regularization-based methods for learning. It also highlights the type of gradient additional terms and neural network implementation methodology.

4.3. Replay-Based Methods

Memory is the core subsystem of the human brain that retains learning over a lifetime. This retention of learning offers the baseline for knowledge reuse and augmentation. This mechanism offers inspiration for the memory-based mechanism for knowledge retention and reuses in ML-based methods to achieve CL [195–198] as shown in Figure 4(b).

Table 4 enlists the recent CL-based methods which adopt regularization-based method for the learning. It also highlights the type of the gradient additional terms and neural network implementation methodology.

4.4. Meta-Learning–Based Methods

Above-mentioned methods may suffer from inductive biases introduced due to the manual setting of parameters. Such biases may cause the suboptimal performance due to differences in objective and expected output [117]. Meta-learning offers better solutions in complex and diverse environments when complicated setting of parameters is required.

Meta-learning is learning to learn. It is inspired by the capability of the human brain that finds new solutions after small learning and experience. By utilizing the learning-to-learn mechanism, this technique tries to learn the dynamic settings as per the working of the brain. Meta-learning is the learning at two levels, i.e., outer loop and inner loop. Outer loop controls the working of inner loops in dynamic scenarios through meta-updates of meta-objectives. The inner loop manages the task-specific fine-tuning and specific task [217]. Generally, meta-learning methods are implemented through support sets and query sets. Support set is utilized for fast learning. The query set is utilized to learn the adaptation [218].

Meta-learning methods can be categorized as model-based, optimization-based, and parameter-based methods [219]. Reference [23] presented a meta-learning methodology to tackle deterministic problems by CL using neural networks. The method utilized a corset as memory and the degree of relationship between global and individual tasks are exploited to tackle forgetting. Ref. [219] implemented meta-learning by the combination of preconditioning matrices to update the gradient for sharing the learned knowledge across multiple tasks to tackle forgetting. Ref. [220] devised an approach in which attention methodology is utilized to empower the meta-learning algorithm to highlight and acquire nontrivial important features, and then, the meta-learning algorithm is enabled to take care of knowledge learned from past tasks and reuse it for future tasks. Ref. [117] proposed a two-level regularization based on meta-learning to regulate the network for knowledge retention and utilization across multiple tasks. The optimization methodology is devised for each task. Ref. [221] exploited the meta-learning to achieve CL through a vision transformer network, i.e., attention to self-attention. A new mask is trained for each new task and added to the main vision-transformed model to achieve learning across multiple tasks. Ref. [222] utilized domain randomization and meta-learning to achieve CL. Meta-learning-based mechanism is used to regularize the loss function related to the learning of multiple domains. The concept of meta-domains/auxiliary domains which are generated by randomized image manipulations is also utilized. Inspired by the slow and fast memory concepts of the human brain, Ref. [218] tried to devise a mechanism of fast and slow updates of weights. In the proposed study, updating weights, feature extraction, and semantic learning mechanisms are defined in fast and slow-based learning.

4.5. Hybrid Methods

Few studies suggest that the above-mentioned approaches may fail even for controlled and well-arranged datasets in scenarios when labels or task IDs are unavailable at the test time [135]. Keeping in view the limitations of the above-mentioned approaches, some studies proposed hybrid solutions that are a combination of the above-mentioned methods [24, 49, 51]. Recent studies show that hybrid methods offer state-of-the-art solutions to avoid catastrophic forgetting particularly for CIL scenarios [25]. Such methods may be categorized in the following types.

5.1. CORe50

Reference [247] proposed CORe50 as a benchmark dataset for object classification. It offers different learning scenarios. It comprised 164,866 images of size 128 × 128. There are a total of fifty types of items which are arranged in ten classes. It offers two types of environments, i.e., indoor and outdoor are covered. Eight scenarios belong to indoor, three scenarios belong to outdoor. These different scenarios mimic the different managed context. Videos and images belong to different lighting, occlusions, pose, and background conditions. Videos and pictures are taken in such a way that all these items are presented to a robot. As an overall, it offers context-based new classes as well as new instances scenarios.

5.2. iCubWorld28

This dataset comprises images of 28 items belonging to seven classes. It represents four different sessions. These sessions represent four different days (days 1–4), in which data are recorded under different environments and conditions. It consists of more than 12,000 images.

5.3. Toys200

It comprised of 200 unique toy-like object samples which are synthetically generated shapes of children toys. Each object has multiple views in the dataset.

5.4. Clear

Reference [250] presented the dataset of different objects that are changed in the shape due to the evolution during the period of 2004–07. These items include cameras, computer, etc. This dataset is constructed from YFCC100M dataset. The main contribution of this dataset is that it portrayed the evolution of real-world objects over the time.

5.5. OpenLORIS

This dataset offers the collection of images taken from RGB-D camera under different environments. It is a candidate dataset for training and testing CL models for domain incremental scenarios. The environment conditions include different occlusions, size, noise, and difficulty levels. It is a real-world dataset. It consists of 186 instances, 63 classes, and 2,138,050 images [251].

5.6. CLAD-C

It is a benchmark dataset for continual object detection [252]. This dataset is constructed based on SODA10M. It comprised of online continuous stream data of 3 days and nights. In the continuous stream, objects appear after every 10 s. The dataset offers data for different domains. The domain shift is exhibited by changing frequencies, time, and weather. It also offers multiple views of the scenes and objects. In order to exactly portray the real-world scenario, the background in the subsequent objects has overlapping.

5.7. CLAD-D

It is also based on SODA10M dataset [252]. It offers a benchmark for autonomous driving. It is proposed for domain incremental learning scenarios. Different domains are portrayed through different highways, day and night time, weather, and location within and outside the city.

5.8. vCLIMB

vCLIMB is video dataset constructed for CL to judge the forgetting in class incremental CL scenarios. It offers a sequence of disjoint tasks with the number of classes which are evenly available over the tasks [131, 159].

5.9. Experimental Configurations Used for CL Methods

There exist various experimental settings which are discussed in the literature related to CL. These settings mimic the real-world situations and natural settings the system has to counter with. Each of these scenarios has its own advantages and challenges [255]. It is pertinent to mention here that there is no agreed standard or guidelines for the formulation of the experimental configurations and division of tasks/problems. In the subsequent subsection, we discuss a few famous experimental configurations reported in the literature related to CL.

6.1. Average Accuracy

6.2. FWT

FWT is to use the learning of known task to expedite learning of future task [263]. This metric measures whether the CL-based system utilizes the knowledge from past task to learn a new task. It needs the availability of evaluation blocks before and after each new task’s first learning to compute the values.

6.3. Average BWT

6.4. Relative Performance (RP)

RP is employed to make the performance of the model equal or improved for the single task. It offers insight into whether a system is showing an improvement over previous attempts to attain CL versus whether a system simply experiencing CL.

6.5. Cumulative Gain

Cumulative gain keeps record of performance after completion of each task during the complete problem solution scenarios [264].

6.6. Performance Maintenance (PM)

PM offers the complete picture of the model with respect to the performance it maintained on each task. It is a relative term which gives the change in performance on the same task when it was attempted for the first time and the performance on the attempts made next time to solve the same task [131, 265]. It compares a model’s performance when it learns a task first time to subsequent times it encounters the task. It does not measure the absolute performance level. It measures the change in performance over complete lifetime of the system [265].

Regarding the values of the PM, if PM > 0, it means the performance on task is getting better over lifetime, if PM = 0, no forgetting; no learning. If PM < 0, it means catastrophic forgetting.

7.1. Image Classification

Tables 5, 6, 7, 8, and 9 show the results of different CL methods for the image classification problem. Table 5 shows the results for CIFAR-10 dataset for CIL and TIL scenarios. It can be observed that as an overall the accuracy for TIL scenarios remains high as compared to CIL scenarios, as CIL scenarios are considered difficult than TIL scenarios. In CIL scenarios, the accuracy of memory replay-based system remained higher as compared to network- and gradient-based methods, whereas the accuracy of architecture-based methods remained at lower side. In TIL scenario, the accuracy of static architecture-based methods remained highest, whereas the accuracy of memory replay method remained lowest. In case, of CIFAR-100 dataset as shown in Table 6, the accuracy of memory-based methods remained at higher side for TIL scenario, whereas the performance of dynamic network method remained at highest end. Table 7 shows the result for ImageNet dataset. For CIL scenario, the accuracy achieved by gradient-based method remained highest. Similarly, for the TIL scenario, the performance of gradient-based method remained highest. Table 10 displays the results of statistical t-test for ImageNet dataset. The table also shows the p-value. t-Statistics and p-value show that Du et al. performed better than other methods. Table 8 shows the results for TinyImageNet dataset for CIL and TIL scenario. It can be observed that the performance of the memory as well as gradient-based methods remained almost same at the higher side. Table 11 displays the results of statistical t-test for Tiny ImageNet dataset. The table also shows the p-value. t-Statistics and p-value show that Cheraghian et al. performed better than other methods. Table 9 shows the results for MNIST dataset for CIL and TIL scenarios. It can be observed that Abati et al. performed better as compared to other methods both in CIL and TIL scenarios.

7.2. Text Classification

Table 12 shows the results for the text classification. It can be observed that the hybrid method (meta + memory-based) outperformed other methods for Yelp Review and Twitter dataset, whereas for the Amazon Review dataset, dynamic network-based method outperformed other methods and its performance remained far better than compared methods. In the case of AG News dataset, the gradient-based method outperformed network and hybrid methods.

7.3. Segmentation

Tables 13 and 14 display the results for the segmentation problem. Table 9 shows the results for Pascal VOC dataset for CIL and TIL scenarios. For CIL scenario, it can be observed that the performance of most of the methods remained in the range of 60–70%. Pseudo-rehearsal memory replay-based method outperformed other compared methods. Similarly, in case of TIL scenario, replay-based methods outperformed other compared methods. It can be observed in Tables 9 and 12 that hybrid methods outperformed other methods in the case of ADE20K dataset for TIL as well as CIL scenarios. Table 15 displays the results of statistical t-test for Pascal VOC dataset. The table also shows the p-value. t-Statistics and p-value show that Douillard et al. performed better than other methods.

7.4. Human Action Recognition

Tables 16 and 17 show the results for human action recognition dataset. It shows that the methods in offline settings outperformed other compared methods for cross-subject (CS) scenarios. It shows results for cross-view (CV) scenario. It can be observed that the performance of top five methods remained almost similar. However, the method in offline settings outperformed other methods. Similar is the case for the NTU RGBD dataset, where the method with offline setting outperformed other methods for CV and CS scenarios.

8.1. Backward Knowledge Transfer

Currently, most approaches addressing catastrophic forgetting in neural networks focus on minimizing the loss of previously learned tasks, ensuring forward knowledge transfer only. However, this approach limits backward knowledge transfer, where insights from new tasks could enhance the performance on earlier tasks. Effective two-way knowledge transfer, both forward and backward, remains a challenging area that requires attention to develop mechanisms that facilitate bidirectional learning.

8.2. Growing Demand of Resources

Most CL-based methods reported in the literature require increasingly substantial resources, such as computing power or storage capacity. For ML systems to emulate the CL capability of the biological brain, they need to operate with limited dedicated resources. Relying on additional resources for learning each new task may not be sustainable or feasible.

8.3. Graceful Forgetting

Graceful forgetting is an essential aspect of CL. Like the biological brain, ML systems need mechanisms to selectively forget nonessential data or outdated knowledge, freeing up resources for critical functions and future learning. Developing effective strategies for graceful forgetting—ones that conserve resources without sacrificing valuable information—remains an active area of research.

8.4. Online Continuous Stream of Data

Task identification and management are crucial features of CL. However, with an online continuous data stream, identifying and managing tasks becomes particularly challenging.

8.5. Biases

Manually setting hyperparameters can introduce biases into the system. Beyond parameter-related biases, the system may also be affected by biases from other factors, such as data imbalances and environmental influences.

8.6. Availability of Real Test Environment

A crucial factor in advancing lifelong learning technology is developing realistic testing environments that specifically assess CL capabilities, moving beyond static, pre-prepared datasets. Currently, most AI models are trained in controlled, well-defined settings. For a model to operate effectively in real-world, dynamic environments, it needs training with data that closely mirror these conditions.

8.7. Multiagent CL

Traditional ML algorithms are typically centralized and rely on single-agent frameworks. However, the rise of the Internet of Things (IoT) has led to unprecedented data growth, creating demand for more sophisticated AI models that operate across multiple agents. Despite this need, multiagent CL remains relatively underexplored compared to traditional AI approaches.

In multiagent reinforcement learning (MARL), transfer learning (TL) is a crucial technique for accelerating learning by facilitating knowledge sharing among agents. TL enables the reuse of knowledge acquired in a source domain, thereby reducing the reliance on extensive datasets and long training times for target tasks. In the context of reinforcement learning (RL), TL can significantly lower the number of samples needed to derive an optimal policy by leveraging previously learned knowledge.

However, applying TL to real-world problems presents three significant challenges. First, most real-world environments are only partially observable, complicating the application of learned knowledge. Second, collecting prior knowledge in unfamiliar domains is inherently difficult. Third, negative transfer—where transferred knowledge hinders rather than aids learning—poses a serious obstacle to progress.

While TL in RL has been the subject of extensive research, a few studies have simultaneously addressed the combined challenges of partial observability, precollecting knowledge, and mitigating negative transfer. Furthermore, the domain of multiagent CL, which involves incorporating both knowledge transfer and incremental learning, remains largely underexplored. This is especially true for scenarios involving evolutionary multitasking, where identifying and transferring relevant and beneficial knowledge across diverse tasks is both essential and highly complex.

We believe that brain biology will continue to serve as a valuable source of inspiration for developing novel lifelong learning approaches. Advances in our understanding of key brain biological mechanisms, such as dynamic memory updating processes like active forgetting, extinction, and memory reconsolidation, will inspire new algorithms beyond those currently discussed. Additionally, deeper insights into intracellular processes like signaling and gene regulation, as well as intercellular communication, could drive lifelong learning innovations beyond the central nervous system.

Lifelong learning systems, with their enhanced capabilities and broader range of behaviors in real-world applications, have the potential to revolutionize fields such as autonomous vehicles, smart cities, and healthcare. Realizing this potential will require ongoing multidisciplinary initiatives that unite researchers from biology, neuroscience, psychology, engineering, and AI. Such collaborations are essential for developing the convergent solutions that this emerging form of AI demands.

Considering the space limitations, this study focuses on the application of CL for computer vision problems using supervised learning methods. Future work may include a comprehensive survey of studies exploring RL and unsupervised learning approaches in domains beyond computer vision. Additionally, other advanced deep learning techniques, such as graph neural networks, could also be examined to provide a broader perspective.