Storage-Efficient Continual Learning

Catastrophic forgetting in sequential deep learning arises from a fixed-capacity architectural constraint: parameter updates encoding new tasks perturb weight subspaces supporting prior representations. We introduce the Knowledge Fusion Network (KFN), which mitigates the stability-plasticity trade-off in class-incremental learning through novelty-gated structural expansion rather than parameter negotiation within a shared capacity budget. KFN decouples task acquisition from knowledge consolidation: each incoming task is learned in an isolated lightweight specialist network (2.8M parameters) and integrated into the global model via a Structural Fusion Module that allocates dedicated channel capacity conditioned on a quantitative novelty score. Dual structural isolation--channel-space separation with temporally phased gradient unfreezing--protects consolidated knowledge. On ten-task Split CIFAR-100, KFN achieves 11.94% minimum task accuracy and 71.16% task-10 accuracy, whereas shared-parameter baselines degrade severely: LwF to 9.07% average accuracy (0.00% min), Replay++ to 8.91% (85.08% forgetting), and Replay to 16.21% (73.44% forgetting); KFN's forgetting (57.92%) is 15.5 pp lower than Replay. Furthermore, KFN prevents logit-scale drift, maintaining near-symmetric task performance. Across CIFAR-100 and EuroSAT satellite imagery, KFN's class-incremental retention (72.7% and 42.5%) substantially exceeds the best baselines (29.6% and 0.94%). Crucially, KFN achieves parameter-efficient linear scaling, adding 25% of the base parameter count per step and requiring 25.1% of naive backbone duplication over ten tasks (28.1M vs 112.0M). With a 37.1% peak GPU memory reduction over full retraining, KFN establishes novelty-gated expansion as a practical path beyond fixed-capacity architectures.

Lifelong robot learning requires acquiring new tasks sequentially without forgetting earlier ones. With large pre-trained vision-language-action (VLA) backbones, adapter-based methods effectively mitigate this challenge. However, their task-specific adapters and selection modules accumulate with the task stream, creating a scalability challenge in on-disk storage and deployment cost under fixed onboard resource budgets. We propose SARAM, a lifelong learning framework that limits storage growth while keeping GPU memory and per-step inference cost independent of the task stream length. For each new task, SARAM reuses frozen adapters from previous tasks and introduces new adapters only at sparsely selected layers, with a duplication penalty encouraging them to capture information not already represented by the adapter library. At deployment, a compact signature derived from backbone activation statistics identifies the task once per episode, and the corresponding task update is merged into the backbone, leaving only the backbone in the control loop. Across four LIBERO suites, SARAM matches or exceeds adapter-based baselines while adding up to 4.5x less on-disk state per task. GPU memory usage and control frequency remain comparable to those of the backbone as the task stream grows. Experiments on a real UR5 arm with a different backbone further demonstrate retention throughout a sequential task stream.

Spiking Neural Networks (SNNs) are promising for energy-efficient neuromorphic computing, but they still suffer from catastrophic forgetting when learning tasks sequentially. This work investigates the integration of Experience Packing and Replay (EPR) as a memory-efficient continual learning strategy for deep SNNs under low-latency constraints. The proposed framework combines Batch Normalization Through Time stabilized VGG9 SNNs with Spike Activation Maps (SAM), which provide forward-only explanations based on spike activity. These explanations are temporally aggregated into spatial attention maps and used to select the most informative image patches for replay, instead of storing full images. Experiments on Split CIFAR10 and Split CIFAR100 show that SAM-guided EPR improves the accuracy-memory trade-off under tight memory budgets, and remains competitive while requiring substantially less storage.

Parameter-efficient continual learning updates a restricted subset of adapter parameters to incorporate sequential tasks while limiting catastrophic forgetting. Gradient attribution is a common rule for selecting that subset, but it requires an additional backward pass over calibration data. We identify a parameter-pool confound in comparisons between initialization-time gradient masks and unrestricted random masks under standard zero-initialized LoRA ($B=0$): the gradient with respect to $A$ is identically zero ($\partial\mathcal{L}/\partial A=0$), so gradient masks select strictly from the $B$ matrix whereas unrestricted random masks draw from both $A$ and $B$. Such a comparison changes both the selection rule and the candidate parameter pool. Isolating the selection criterion by restricting both approaches to the same parameter pool, we find a negative result: a uniform random mask consistently achieves lower loss-based backward transfer than gradient attribution across seven seeds on Qwen2.5-3B. The effect is concentrated on the medical task, which contributes 80% of the aggregate; code and math move in the same direction, and legal shows no reliable difference. This advantage is consistent with a stability--plasticity trade-off under the tested task sequence, in which random masking slightly reduces single-task fitting. On the two discrete-option tasks (Medical and Legal), the mean accuracy-BWT difference is not statistically distinguishable from zero.

Perception for embodied agents is video-based, often multi-view (ego, exo, or both), and inherently continual, with simultaneous task and viewpoint shifts. Yet continual learning (CL) remains dominated by exo-only recognition tasks, obscuring behavior under these real-world coupled shifts. We introduce Continual Ego, E}xo, and Ego-Exo Learning (CE$^4$L), a unified multi-view CL benchmark spanning four representative tasks: cross-view referenced skill assessment, temporal action segmentation, cross-view association, and action anticipation & planning. CE$^4$L highlights challenges largely absent in prior CL benchmarks, including cross-view correspondence, view-dependent asynchrony, and heterogeneous semantic objectives. To this end, we propose Video Incremental Subspace-routed Task Adapters (VISTA), a parameter-efficient baseline method that stores task-specific updates in lightweight adapters and performs training-free routing via residual distance to task-specific whitened subspaces estimated from second-order statistics. Extensive experiments demonstrate the significantly varied efficacy of representative CL methods across CE$^4$L settings, while VISTA is consistently competitive and achieves state-of-the-art overall performance. Our source code for benchmarks and methods is available at https://github.com/AnAppleCore/CE4L .

We introduce Neural Subspace Reallocation (NSR), which reframes continual learning as memory management over parameter subspaces. Instead of treating Low-Rank Adaptation (LoRA) modules as disposable per-task adapters, NSR manages them as compressible, retrievable memory units on a frozen backbone through a recurring cycle: (1) compress learned LoRAs via SVD, (2) reserve them in a TaskKnowledgeBank, (3) recall related past LoRAs by embedding similarity to warm-start new or returning tasks, and (4) reallocate the active subspace accordingly, with distillation protecting prior tasks. We prove that in cyclic environments any memoryless allocation policy incurs cumulative regret Omega(T(M-1)Delta_switch) relative to a history-aware policy backed by the Bank (Theorem 1). Empirically, on Split-CIFAR-100 the Bank reduces cyclic recovery time by 10x, exactly as predicted, and on the heterogeneous 5-Datasets benchmark NSR achieves the highest accuracy and the least forgetting, about 9x closer to zero backward transfer than the memoryless heuristics. Crucially, we run a controlled study that isolates which component matters: holding the Bank fixed and varying only the allocation rule, we find that a simple similarity-based retrieval rule matches or beats a learned reinforcement-learning controller (recovering recurring tasks in 0 vs 1.8 steps and reaching equal accuracy). Our central, honest finding is therefore that the memory mechanism -- compression and similarity retrieval -- rather than a learned allocation policy, drives continual-learning performance under fixed capacity. A memory-budget analysis confirms the compressed Bank stays small -- 0.29 MB of parameter memory per task -- so a top-K retention cap bounds the total footprint while preserving fast recovery for retained tasks.

While prompt-based parameter-efficient continual learning mitigates catastrophic forgetting by isolating task-specific prompts, this isolation also limits later tasks from improving earlier ones, leaving backward knowledge transfer underexplored. We address this limitation by proposing Selective bAckward refinement for positive Backward knowledge transfER (SABER), a replay-free framework that enables controlled backward transfer in prompt-based continual learning. SABER determines when backward refinement is beneficial using complementary task-correlation criteria based on prompt-gradient geometry and loss-distribution similarity, and how to perform refinement safely by restricting updates to non-interfering directions in the prompt parameter space. Extensive experiments across multiple continual learning benchmarks and diverse pretrained backbones, including T5-Large, LLaMA, and Qwen, demonstrate that SABER consistently achieves positive backward transfer while maintaining strong overall average performance. Code is available at https://github.com/OptMN-Lab/SABER-ICML-2026/.

Model merging provides a compelling paradigm for integrating specialized expertise into a unified multi-task model, a goal that aligns naturally with the sequential knowledge acquisition in continual learning (CL). However, the requirement for preserving diverse forms of previous knowledge conflicts with the storage limitations inherent to CL. In this paper, we systematically analyze existing model merging methods under the constraints of CL. We find that current methods prioritize global alignment, which often leads to the accumulation and amplification of task-specific errors within the continuous data stream; and the vanishing gradients at the onset of subsequent tasks frequently cause optimization to stagnate. These leave the merged model in a suboptimal state at the beginning of the next training phase. To address these challenges, we propose Trajectory Regularized Merging (TRM), a framework that reformulates the merging phase as an optimization process within an augmented trajectory subspace. Our framework integrates three synergistic objectives including task alignment, prediction consistency, and gradient responsiveness to concurrently preserve merged model's historical stability and re-activate optimization dynamics. Extensive experimental results demonstrate that our method achieves state-of-the-art performance across multiple benchmarks.

An agent that operates sequentially must incorporate new experience without forgetting old experience, under a fixed memory budget. We propose a framework in which memory is not a parameter vector but a stochastic process: a Bridge Diffusion on a replay interval $[0,1]$, whose terminal marginal encodes the present and whose intermediate marginals encode the past. New experience is incorporated via a three-step Compress--Add--Smooth (CAS) recursion. We test the framework on the class of models with marginal probability densities modeled via Gaussian mixtures of fixed number of components~$K$ in $d$ dimensions; temporal complexity is controlled by a fixed number~$L$ of piecewise-linear protocol segments whose nodes store Gaussian-mixture states. The entire recursion costs $O(LKd^2)$ flops per day -- no backpropagation, no stored data, no neural networks -- making it viable for controller-light hardware. Forgetting in this framework arises not from parameter interference but from lossy temporal compression: the re-approximation of a finer protocol by a coarser one under a fixed segment budget. We find that the retention half-life scales linearly as $a_{1/2}\approx c\,L$ with a constant $c>1$ that depends on the dynamics but not on the mixture complexity~$K$, the dimension~$d$, or the geometry of the target family. The constant~$c$ admits an information-theoretic interpretation analogous to the Shannon channel capacity. The stochastic process underlying the bridge provides temporally coherent movie'' replay -- compressed narratives of the agent's history, demonstrated visually on an MNIST latent-space illustration. The framework provides a fully analytical Ising model'' of continual learning in which the mechanism, rate, and form of forgetting can be studied with mathematical precision.

The continual learning literature has rapidly shifted from traditional class incremental learning (CIL) techniques to foundation model (FM)-based CIL methods without a clear understanding of how these newer approaches compare to strong, lightweight convolutional baselines. This abrupt transition has created a substantial methodological gap, making it difficult to assess whether recent FM-based CIL progress reflects genuine advances or merely the absence of rigorous baselines. To address this gap, we introduce Pruned Adaptation Modules (PAM), a simple yet effective method that freezes the vast majority of the pre-trained ResNet while enabling scalable continual adaptation through sparse task-specific layers. PAM yields up to a ~5x reduction in trainable parameters and a ~6x reduction in total parameters, significantly reducing the cost of continual updates. Across diverse benchmarks, PAM consistently mitigates catastrophic forgetting and outperforms state-of-the-art FM-based CIL approaches. Our findings position PAM as a strong and transparent baseline that helps bridge the gap between traditional and FM-based CIL, guiding future research for a more accurate assessment of true progress in continual adaptation. The code can be found at: https://github.com/ElifCerenGokYildirim/PAM.

Class Incremental Learning (CIL) poses a fundamental challenge: maintaining a balance between the plasticity required to learn new tasks and the stability needed to prevent catastrophic forgetting. While expansion-based methods effectively mitigate forgetting by adding task-specific parameters, they suffer from uncontrolled architectural growth and memory overhead. In this paper, we propose a novel dynamic scaling framework that adaptively manages model capacity through a cyclic "GRow, Assess, ComprEss" (GRACE) strategy. Crucially, we supplement backbone expansion with a novel saturation assessment phase that evaluates the utilization of the model's capacity. This assessment allows the framework to make informed decisions to either expand the architecture or compress the backbones into a streamlined representation, preventing parameter explosion. Experimental results demonstrate that our approach achieves state-of-the-art performance across multiple CIL benchmarks, while reducing memory footprint by up to a 73% compared to purely expansionist models.