Search Results for « Catastrophic Forgetting »

Continual learning is a fundamental challenge in artificial intelligence (AI) that aims to enable models to learn from a continuous stream of data while retaining previously acquired knowledge. Unlike traditional machine learning, which operates in static environments, continual learning requires algorithms to adapt incrementally to new tasks and evolving data distributions without catastrophic forgetting. This capability is crucial for deploying AI systems in dynamic real-world applications, including robotics, healthcare, natural language processing, and cybersecurity. This survey provides a comprehensive overview of continual learning, covering its core principles, learning paradigms, and major approaches. We discuss key strategies for mitigating catastrophic forgetting, including replay-based methods, regularization techniques, and dynamic architectures. Additionally, we explore diverse application domains where continual learning plays a crucial role, emphasizing its significance in lifelong learning scenarios. Despite recent advancements, several challenges remain, such as scalability constraints, the lack of standardized benchmarks, and the need for biologically inspired learning mechanisms. We outline open research directions, including memory-efficient learning, adaptive self-supervised techniques, and fairness-aware continual learning. By addressing these challenges, the AI community can develop more robust and flexible models capable of learning continuously over time. Through this survey, we aim to provide a structured foundation for researchers and practitioners interested in continual learning, highlighting its potential to drive the next generation of adaptive AI systems. We discuss the broader impact of continual learning on artificial intelligence and outline future directions to advance the field further.

Recent advancements in Large Language Models (LLMs) have showcased their remarkable capabilities in text understanding and generation. However, even stronger LLMs are susceptible to acquiring erroneous or obsolete information from the training corpus. Direct secondary fine-tuning with data containing new knowledge may be ineffective in updating knowledge due to the conflict between old and new knowledge. In this paper, we propose a new paradigm for fine-tuning called DFT (Delicate Fine-Tuning ).This method utilizes parametric arithmetic to precisely pinpoint the location of knowledge and update only the minimal set of relevant parameters . Experimental results on two publicly available datasets demonstrate that our proposed DFT can obviously improve the knowledge updating performance of full fine-tuning , simultaneously outperforming the existing baselines in most cases.

Recent advancements in Large Language Models (LLMs) have showcased their remarkable capabilities in text understanding and generation. However, even stronger LLMs are susceptible to acquiring erroneous or obsolete information from the training corpus. Direct secondary fine-tuning with data containing new knowledge may be ineffective in updating knowledge due to the conflict between old and new knowledge. In this paper, we propose a new paradigm for fine-tuning called DFT (Delicate Fine-Tuning).This method utilizes parametric arithmetic to precisely pinpoint the location of knowledge and update only the minimal set of relevant parameters. Experimental results on two publicly available datasets demonstrate that our proposed DFT can obviously improve the knowledge updating performance of full fine-tuning , simultaneously outperforming the existing baselines in most cases.

The hippocampus is the brain's central locus for memory processing. In a widely accepted hypothesis, the hippocampus stores short-term memories in attractor states, which are then consolidated in the neocortex as long-term memories. Hippocampal replay activities play a pivotal role in memory consolidation, but memories are only transiently, not persistently, replayed, casting doubt on the attractor memory hypothesis. Concepts like memory capacity, the upper bound on the number of storable stable memories in a neural network, may also need to be revised if their offline transient activation is more essential for memory consolidation. Here, we explore the biologically plausible mechanism of offline memory processing by asking a recurrent spiking network to replay reliably as many stored assemblies as possible. We demonstrate the importance of memory-assembly-specific inhibitory plasticity for replaying diverse memory assemblies. Furthermore, our results highlight the role of transient memory states in mitigating catastrophic forgetting.

Abstract It is widely assumed that continued pretraining (CPT) on a narrow, out-of-distribution corpus such as raw biological sequence must trade away a general-purpose model’s broad competence — the “alignment tax” or catastrophic-forgetting intuition. We test this directly, without any new training, by re-analyzing three checkpoints from a single lineage of a 26B-parameter Mixture-of-Experts model (Gemma-4-26B-A4B): the instruction-tuned base, the same model after biological CPT (8.7B tokens of DNA, protein, and biomedical text), and after subsequent supervised fine-tuning (SFT). Across three independent capability axes — general knowledge/reasoning (MMLU, ARC, HellaSwag), code generation (MBPP), and biomedical knowledge (BixBench) — we find that biological CPT does not degrade the model; it lifts it: MMLU +13 points, MBPP pass@1 nearly doubles (0.33 →0.63), and BixBench discrimination rises sharply (MCC 0.23 → 0.92). The single measured regression is truthfulness (TruthfulQA 8.8 points), a small and interpretable domain drift. A clean vocabulary-expansion ablation ( < 0.4 pt on every general metric) confirms the gains are attributable to CPT, not tokenizer changes. Crucially, subsequent SFT narrows the model back: all three axes fall to near-base levels, revealing a consistent division of labor — CPT re-organizes and lifts the shared capability substrate; SFT cashes it out onto target tasks . We argue this reframes biological sequence not as a competitor for a foundation model’s capacity but as a form of structured scientific data that reshapes its capability profile, and that CPT and SFT should be budgeted as complementary rather than substitutable stages. All checkpoints, evaluation code, and per-example outputs are public. Highlights A training-free re-analysis of one 26B MoE lineage isolates the effect of biological continued pretraining (CPT) from tokenizer changes and from fine-tuning. Biological CPT does not cause catastrophic forgetting; it raises general knowledge (MMLU +13 pts) and code generation (MBPP pass@1 0.33 → 0.63). CPT also makes chain-of-thought reasoning 41% shorter and near-backtrack-free while pre-serving accuracy — an effect invisible to accuracy metrics. A consistent CPT-lifts / SFT-narrows division of labor recurs across four axes, reframing biological sequence as structured scientific data that reshapes a model’s capability profile. The Bigger Picture Adapting a general-purpose AI model to a specialized domain — here, the language of DNA and proteins — is usually assumed to come at a cost: teach it biology and it forgets how to reason about everything else. This “no free lunch” intuition shapes how practitioners budget compute and whether they attempt domain adaptation at all. We test the assumption directly, and without running any new training, by comparing three snapshots of the same model taken before and after biological training. The result overturns the intuition: feeding the model raw biological sequence made it better at general knowledge, at writing code, and even changed how it reasons — producing shorter, more decisive chains of thought without losing accuracy. The gains appear during the sequence-pretraining stage and are partly given back during task-specific fine-tuning, revealing that the two stages play complementary rather than interchangeable roles. This suggests a broader principle for data-centric AI: structured scientific data — biological sequence today, and by extension code, mathematics, and chemistry — is not merely knowledge to be absorbed but a lever that reshapes what a foundation model can do.

Catastrophic forgetting, the tendency of learning systems to lose previously acquired knowledge when trained on new information, remains one of the most fundamental challenges in continual learning. This dissertation contributes to a deeper understanding of why catastrophic forgetting occurs and investigates the mechanisms that enable learning models to maintain stability while acquiring new knowledge over time. The core of this work centers on the Cobweb framework, a psychologically inspired concept formation model that incrementally builds a hierarchy of concepts. A novel extension of this framework, Cobweb/4V, is introduced and systematically evaluated for its capacity to learn visual concepts in a continual manner without suffering from catastrophic forgetting. Motivated by the observed robustness of Cobweb/4V under continual learning, this research examines three hypotheses regarding factors that may contribute to such stability: (1) the adaptive structure hypothesis, which suggests that the ability to dynamically restructure the concept hierarchy as new data arrive supports continual learning by flexibly allocating representational capacity; (2) the sparse update hypothesis, which suggests that restricting parameter updates to localized regions of the model, rather than applying global changes, helps preserve prior knowledge by reducing representational overlap; and (3) the information-theoretic hypothesis, which proposes that employing closed-form updates based on sufficiency statistics enables precise, incremental learning without revisiting past data, offering potential advantages over gradient-based methods when previous data are unavailable. These hypotheses are evaluated through comparative experiments on datasets of varying complexity. By combining cognitive inspiration with controlled empirical evaluations, this dissertation offers new insights into the dynamics of memory retention and interference. These contributions not only clarify the foundations of catastrophic forgetting, but also point toward promising directions for developing continual learning systems that are both stable and adaptable.

Catastrophic forgetting is a fundamental challenge in sequential learning for artificial neural networks, particularly spiking neural networks (SNNs), which aim to emulate biologically plausible neuronal dynamics. In this study, we investigate the role of sleep-mediated replay in mitigating catastrophic forgetting during sequential task learning. We trained a multilayer SNN on two sequentially presented image classification tasks (Task A and Task B) using spike-timing-dependent plasticity (STDP) learning rules. During offline periods, we implemented spontaneous replay of Task A patterns to emulate biological sleep consolidation. Our results demonstrate that sleep-mediated replay keeps Task A performance (0.589 before Task B, 0.206 after sleep) while allowing effective acquisition of Task B (0.672 before sleep). In contrast, networks trained sequentially without sleep exhibit severe forgetting (Task A drops to 0.011 after Task B). These findings suggest that sleep-like replay mechanisms can enable continual learning in SNNs, providing a computationally and biologically grounded approach to memory consolidation. We further analyze synaptic weight distributions and firing rate dynamics, revealing that sleep stabilizes task-specific synaptic patterns without impeding new learning. This work bridges computational neuroscience and artificial intelligence, suggesting that sleep-like replay could be an effective approach for continual learning in SNNs.

Abstract We introduce LatentShift, a replay-free continual learning method built on a simple linear-algebraic construction: after each task, an SVD of the network's bottleneck activations identifies an orthonormal basis for the occupied latent subspace, this basis is appended to a cumulative archive, and all subsequent gradients flowing through the bottleneck are projected onto the orthogonal free subspace. The construction yields three analytic results: first-order latent-subspace protection—to first order in the learning rate, projected updates produce no change in archived directions; a capacity theorem T_max = floor(d/r-bar) relating tasks to latent dimension d and average per-task rank r-bar; and an explicit second-order bound on residual drift from encoder non-linearity. The first-order guarantee operates in latent space; its parameter-space implication is approximate and depends on a Jacobian-commutation assumption that we make explicit. Across six benchmarks—including a 50-task stress-test sequence—two architectures (ResNet-18 and ViT-Tiny), and thirteen competing approaches, LatentShift consistently attains among the lowest forgetting and roughly halves the forgetting of the strongest replay-based baseline on CIFAR-100 and TinyImageNet, while maintaining competitive accuracy at substantially reduced computational and storage cost.