arXiv:2511.17439v1 Announce Type: new 
Abstract: Continual learning aims to enable neural networks to acquire new knowledge without forgetting previously learned information. While recent prompt-based methods perform strongly in class-incremental settings, they remain vulnerable under domain shifts, where the input distribution changes but the label space remains fixed. This exposes a persistent problem known as representation drift. Shared representations evolve in ways that overwrite previously useful features and cause forgetting even when prompts isolate task-specific parameters. To address this issue, we introduce InTAct, a method that preserves functional behavior in shared layers without freezing parameters or storing past data. InTAct captures the characteristic activation ranges associated with previously learned tasks and constrains updates to ensure the network remains consistent within these regions, while still allowing for flexible adaptation elsewhere. In doing so, InTAct stabilizes the functional role of important neurons rather than directly restricting parameter values. The approach is architecture-agnostic and integrates seamlessly into existing prompt-based continual learning frameworks. By regulating representation changes where past knowledge is encoded, InTAct achieves a principled balance between stability and plasticity. Across diverse domain-incremental benchmarks, including DomainNet and ImageNet-R, InTAct consistently reduces representation drift and improves performance, increasing Average Accuracy by up to 8 percentage points over state-of-the-art baselines.

تم تقديم InTAct، وهي طريقة جديدة للتعلم المستمر، لمعالجة تحدي انحراف التمثيل في الشبكات العصبية. تتيح هذه الطريقة للشبكات اكتساب معارف جديدة مع الحفاظ على المعلومات التي تم تعلمها سابقًا، خاصة في السيناريوهات التي تحدث فيها تغييرات في المجال ولكن تظل مساحة التسميات ثابتة.

Se ha introducido InTAct, un nuevo método para el aprendizaje continuo, para abordar el desafío de la deriva de representación en las redes neuronales. Este método permite que las redes adquieran nuevos conocimientos mientras preservan la información previamente aprendida, especialmente en escenarios donde ocurren cambios de dominio pero el espacio de etiquetas permanece inalterado.

InTAct, une nouvelle méthode pour l'apprentissage continu, a été introduite pour relever le défi de la dérive de représentation dans les réseaux de neurones. Cette méthode permet aux réseaux d'acquérir de nouvelles connaissances tout en préservant les informations précédemment apprises, en particulier dans les scénarios où des changements de domaine se produisent mais où l'espace d'étiquettes reste inchangé.

InTAct, a new method for continual learning, has been introduced to address the challenge of representation drift in neural networks. This method allows networks to acquire new knowledge while preserving previously learned information, particularly in scenarios where domain shifts occur but the label space remains unchanged.

InTAct: Interval-based Task Activation Consolidation for Continual Learning

arXiv:2505.15241v2 Announce Type: replace 
Abstract: Despite progress in geometry-aware domain adaptation, current methods such as GAMA still suffer from two unresolved issues: (1) insufficient disentanglement of task-relevant and task-irrelevant manifold dimensions, and (2) rigid perturbation schemes that ignore per-class alignment asymmetries. To address this, we propose GAMA++, a novel framework that introduces (i) latent space disentanglement to isolate label-consistent manifold directions from nuisance factors, and (ii) an adaptive contrastive perturbation strategy that tailors both on- and off-manifold exploration to class-specific manifold curvature and alignment discrepancy. We further propose a cross-domain contrastive consistency loss that encourages local semantic clusters to align while preserving intra-domain diversity. Our method achieves state-of-the-art results on DomainNet, Office-Home, and VisDA benchmarks under both standard and few-shot settings, with notable improvements in class-level alignment fidelity and boundary robustness. GAMA++ sets a new standard for semantic geometry alignment in transfer learning.

تم تقديم إطار جديد يسمى GAMA++ لتحسين التكيف مع المجال الواعي للهندسة، حيث يعالج مشكلات التفكيك وأنظمة الاضطراب الصارمة التي تؤثر على الأداء. تستخدم هذه الطريقة التفكيك في الفضاء الكامن واستراتيجية الاضطراب التبايني التكيفية المخصصة لاحتياجات الفئات المحددة، محققة نتائج رائدة على معايير مثل DomainNet وOffice-Home وVisDA.

Se ha introducido un nuevo marco llamado GAMA++ para mejorar la adaptación de dominio consciente de la geometría, abordando problemas de desentrelazado y esquemas de perturbación rígidos que afectan el rendimiento. Este método emplea desentrelazado del espacio latente y una estrategia de perturbación contrastiva adaptativa adaptada a las necesidades específicas de cada clase, logrando resultados de vanguardia en benchmarks como DomainNet, Office-Home y VisDA.

Un nouveau cadre nommé GAMA++ a été introduit pour améliorer l'adaptation de domaine consciente de la géométrie, en abordant les problèmes de désentrelacement et de schémas de perturbation rigides qui affectent les performances. Cette méthode utilise le désentrelacement de l'espace latent et une stratégie de perturbation contrastive adaptative adaptée aux besoins spécifiques des classes, atteignant des résultats à la pointe sur des références telles que DomainNet, Office-Home et VisDA.

A novel framework named GAMA++ has been introduced to enhance geometry-aware domain adaptation, addressing issues of disentanglement and rigid perturbation schemes that affect performance. This method employs latent space disentanglement and an adaptive contrastive perturbation strategy tailored to class-specific needs, achieving state-of-the-art results on benchmarks like DomainNet, Office-Home, and VisDA.

Disentangled Geometric Alignment with Adaptive Contrastive Perturbation for Reliable Domain Transfer

arXiv:2505.15191v2 Announce Type: replace 
Abstract: Transfer learning under domain shift remains a fundamental challenge due to the divergence between source and target data manifolds. In this paper, we propose MAADA (Manifold-Aware Adversarial Data Augmentation), a novel framework that decomposes adversarial perturbations into on-manifold and off-manifold components to simultaneously capture semantic variation and model brittleness. We theoretically demonstrate that enforcing on-manifold consistency reduces hypothesis complexity and improves generalization, while off-manifold regularization smooths decision boundaries in low-density regions. Moreover, we introduce a geometry-aware alignment loss that minimizes geodesic discrepancy between source and target manifolds. Experiments on DomainNet, VisDA, and Office-Home show that MAADA consistently outperforms existing adversarial and adaptation methods in both unsupervised and few-shot settings, demonstrating superior structural robustness and cross-domain generalization.

تم تقديم إطار جديد يسمى MAADA (تعزيز البيانات العدائية الواعي بالمنحنى) لمعالجة تحديات التعلم الانتقالي تحت تغير المجال، من خلال تفكيك الاضطرابات العدائية إلى مكونات على المنحنى وخارج المنحنى. يعزز هذا النهج من قوة النموذج والتعميم من خلال تقليل الفجوات الجيوديسية بين منحنيات البيانات المصدر والهدف، كما يتضح من التجارب على DomainNet وVisDA وOffice-Home.

Se ha introducido un nuevo marco llamado MAADA (Aumento de Datos Adversariales Consciente del Manifold) para abordar los desafíos del aprendizaje por transferencia ante el cambio de dominio, descomponiendo eficazmente las perturbaciones adversariales en componentes en-manifold y fuera-manifold. Este enfoque mejora la robustez del modelo y la generalización al minimizar las discrepancias geodésicas entre los manifolds de datos fuente y objetivo, como se demuestra en experimentos realizados en DomainNet, VisDA y Office-Home.

Un nouveau cadre nommé MAADA (Manifold-Aware Adversarial Data Augmentation) a été introduit pour relever les défis de l'apprentissage par transfert face au changement de domaine, décomposant efficacement les perturbations adversariales en composants sur-manifold et hors-manifold. Cette approche améliore la robustesse et la généralisation du modèle en minimisant les écarts géodésiques entre les manifolds de données source et cible, comme le montrent les expériences sur DomainNet, VisDA et Office-Home.

A novel framework named MAADA (Manifold-Aware Adversarial Data Augmentation) has been introduced to tackle the challenges of transfer learning under domain shift, effectively decomposing adversarial perturbations into on-manifold and off-manifold components. This approach enhances model robustness and generalization by minimizing geodesic discrepancies between source and target data manifolds, as demonstrated through experiments on DomainNet, VisDA, and Office-Home.

InTAct: Interval-based Task Activation Consolidation for Continual Learning

Was this article worth reading? Share it

One More Thing in AI

AIPortalX

Dyad