Large language models (LLMs) are demonstrably capable of cross-lingual transfer, but can produce inconsistent output when prompted with the same queries written in different languages. To understand how language models are able to generalize knowledge from one language to the others, we apply the logit lens to interpret the implicit steps taken by LLMs to solve multilingual multi-choice reasoning questions. We find LLMs predict inconsistently and are less accurate because they rely on subspaces of individual languages, rather than working in a shared semantic space. While larger models are more multilingual, we show their hidden states are more likely to dissociate from the shared representation compared to smaller models, but are nevertheless more capable of retrieving knowledge embedded across different languages. Finally, we demonstrate that knowledge sharing can be modulated by steering the models' latent processing towards the shared semantic space. We find reinforcing utilization of the shared space improves the models' multilingual reasoning performance, as a result of more knowledge transfer from, and better output consistency with English.
Outside of transfer learning settings, reinforcement learning agents start their learning process from a clean slate. As a result, such agents have to go through a slow process to learn even the most obvious skills required to solve a problem. In this paper, we present InnateCoder, a system that leverages human knowledge encoded in foundation models to provide programmatic policies that encode "innate skills" in the form of temporally extended actions, or options. In contrast to existing approaches to learning options, InnateCoder learns them from the general human knowledge encoded in foundation models in a zero-shot setting, and not from the knowledge the agent gains by interacting with the environment. Then, InnateCoder searches for a programmatic policy by combining the programs encoding these options into larger and more complex programs. We hypothesized that InnateCoder's way of learning and using options could improve the sampling efficiency of current methods for learning programmatic policies. Empirical results in MicroRTS and Karel the Robot support our hypothesis, since they show that InnateCoder is more sample efficient than versions of the system that do not use options or learn them from experience.
Federated Parameter-Efficient Fine-Tuning (FedPEFT) reduces communication and computation costs in federated fine-tuning of pre-trained models by updating only a small subset of model parameters. However, existing approaches assume static data distributions, failing to adequately address real-world scenarios where new classes continually emerge, particularly in Federated Class Incremental Learning (FCIL). FCIL faces two key challenges: catastrophic forgetting and performance degradation caused by non-IID data across clients. Unlike current methods that maintain separate task-specific components or suffer from aggregation noise during parameter aggregation, we propose Federated Task-agnostic Low-rank Residual Adaptation (Fed-TaLoRA), a novel parameter-efficient approach for fine-tuning in resource-constrained FCIL scenarios. Specifically, we fine-tune only shared task-agnostic LoRA parameters across sequential tasks, effectively mitigating catastrophic forgetting while enabling efficient knowledge transfer among clients. Based on a theoretical analysis of aggregation, we develop a novel residual weight update mechanism that ensures accurate knowledge consolidation with minimal overhead. Our methodological innovations are attributed to three key strategies: task-agnostic adaptation, post-aggregation model calibration, and strategic placement of LoRA modules. Extensive experiments on multiple benchmark datasets demonstrate that Fed-TaLoRA consistently outperforms state-of-the-art methods in diverse data heterogeneity scenarios while substantially reducing resource requirements.
This work introduces CLIP-aware Domain-Adaptive Super-Resolution (CDASR), a novel framework that addresses the critical challenge of domain generalization in single image super-resolution. By leveraging the semantic capabilities of CLIP (Contrastive Language-Image Pre-training), CDASR achieves unprecedented performance across diverse domains and extreme scaling factors. The proposed method integrates CLIP-guided feature alignment mechanism with a meta-learning inspired few-shot adaptation strategy, enabling efficient knowledge transfer and rapid adaptation to target domains. A custom domain-adaptive module processes CLIP features alongside super-resolution features through a multi-stage transformation process, including CLIP feature processing, spatial feature generation, and feature fusion. This intricate process ensures effective incorporation of semantic information into the super-resolution pipeline. Additionally, CDASR employs a multi-component loss function that combines pixel-wise reconstruction, perceptual similarity, and semantic consistency. Extensive experiments on benchmark datasets demonstrate CDASR's superiority, particularly in challenging scenarios. On the Urban100 dataset at $\times$8 scaling, CDASR achieves a significant PSNR gain of 0.15dB over existing methods, with even larger improvements of up to 0.30dB observed at $\times$16 scaling.
Knowledge distillation (KD) is a widely adopted technique for transferring knowledge from a high-capacity teacher model to a smaller student model by aligning their output distributions. However, existing methods often underperform in fine-grained visual recognition tasks, where distinguishing subtle differences between visually similar classes is essential. This performance gap stems from the fact that conventional approaches treat the teacher's output logits as a single, undifferentiated signal-assuming all contained information is equally beneficial to the student. Consequently, student models may become overloaded with redundant signals and fail to capture the teacher's nuanced decision boundaries. To address this issue, we propose Fine-Grained Knowledge Distillation (FiGKD), a novel frequency-aware framework that decomposes a model's logits into low-frequency (content) and high-frequency (detail) components using the discrete wavelet transform (DWT). FiGKD selectively transfers only the high-frequency components, which encode the teacher's semantic decision patterns, while discarding redundant low-frequency content already conveyed through ground-truth supervision. Our approach is simple, architecture-agnostic, and requires no access to intermediate feature maps. Extensive experiments on CIFAR-100, TinyImageNet, and multiple fine-grained recognition benchmarks show that FiGKD consistently outperforms state-of-the-art logit-based and feature-based distillation methods across a variety of teacher-student configurations. These findings confirm that frequency-aware logit decomposition enables more efficient and effective knowledge transfer, particularly in resource-constrained settings.
Recent State Space Models (SSM), especially Mamba, have demonstrated impressive performance in visual modeling and possess superior model efficiency. However, the application of Mamba to visual tasks suffers inferior performance due to three main constraints existing in the sequential model: 1) Casual computing is incapable of accessing global context; 2) Long-range forgetting when computing the current hidden states; 3) Weak spatial structural modeling due to the transformed sequential input. To address these issues, we investigate a simple yet powerful vision task Adaptor for Mamba models, which consists of two functional modules: Adaptor-T and Adaptor-S. When solving the hidden states for SSM, we apply a lightweight prediction module Adaptor-T to select a set of learnable locations as memory augmentations to ease long-range forgetting issues. Moreover, we leverage Adapator-S, composed of multi-scale dilated convolutional kernels, to enhance the spatial modeling and introduce the image inductive bias into the feature output. Both modules can enlarge the context modeling in casual computing, as the output is enhanced by the inaccessible features. We explore three usages of Mamba-Adaptor: A general visual backbone for various vision tasks; A booster module to raise the performance of pretrained backbones; A highly efficient fine-tuning module that adapts the base model for transfer learning tasks. Extensive experiments verify the effectiveness of Mamba-Adaptor in three settings. Notably, our Mamba-Adaptor achieves state-of the-art performance on the ImageNet and COCO benchmarks.
Transfer learning across domains with distribution shift remains a fundamental challenge in building robust and adaptable machine learning systems. While adversarial perturbations are traditionally viewed as threats that expose model vulnerabilities, recent studies suggest that they can also serve as constructive tools for data augmentation. In this work, we systematically investigate the role of adversarial data augmentation (ADA) in enhancing both robustness and adaptivity in transfer learning settings. We analyze how adversarial examples, when used strategically during training, improve domain generalization by enriching decision boundaries and reducing overfitting to source-domain-specific features. We further propose a unified framework that integrates ADA with consistency regularization and domain-invariant representation learning. Extensive experiments across multiple benchmark datasets -- including VisDA, DomainNet, and Office-Home -- demonstrate that our method consistently improves target-domain performance under both unsupervised and few-shot domain adaptation settings. Our results highlight a constructive perspective of adversarial learning, transforming perturbation from a destructive attack into a regularizing force for cross-domain transferability.
Training high-performing Small Language Models (SLMs) remains costly, even with knowledge distillation and pruning from larger teacher models. Existing work often faces three key challenges: (1) information loss from hard pruning, (2) inefficient alignment of representations, and (3) underutilization of informative activations, particularly from Feed-Forward Networks (FFNs). To address these challenges, we introduce Low-Rank Clone (LRC), an efficient pre-training method that constructs SLMs aspiring to behavioral equivalence with strong teacher models. LRC trains a set of low-rank projection matrices that jointly enable soft pruning by compressing teacher weights, and activation clone by aligning student activations, including FFN signals, with those of the teacher. This unified design maximizes knowledge transfer while removing the need for explicit alignment modules. Extensive experiments with open-source teachers (e.g., Llama-3.2-3B-Instruct, Qwen2.5-3B/7B-Instruct) show that LRC matches or surpasses state-of-the-art models trained on trillions of tokens--while using only 20B tokens, achieving over 1,000x training efficiency. Our codes and model checkpoints are available at https://github.com/CURRENTF/LowRankClone and https://huggingface.co/collections/JitaiHao/low-rank-clone-lrc-6828389e96a93f1d4219dfaf.
One of the major challenges in the field of computer vision especially for detection, segmentation, recognition, monitoring, and automated solutions, is the quality of images. Image degradation, often caused by factors such as rain, fog, lighting, etc., has a negative impact on automated decision-making.Furthermore, several image restoration solutions exist, including restoration models for single degradation and restoration models for multiple degradations. However, these solutions are not suitable for real-time processing. In this study, the aim was to develop a real-time image restoration solution for video surveillance. To achieve this, using transfer learning with ResNet_50, we developed a model for automatically identifying the types of degradation present in an image to reference the necessary treatment(s) for image restoration. Our solution has the advantage of being flexible and scalable.
Transferring linguistic knowledge from a pretrained language model (PLM) to acoustic feature learning has proven effective in enhancing end-to-end automatic speech recognition (E2E-ASR). However, aligning representations between linguistic and acoustic modalities remains a challenge due to inherent modality gaps. Optimal transport (OT) has shown promise in mitigating these gaps by minimizing the Wasserstein distance (WD) between linguistic and acoustic feature distributions. However, previous OT-based methods overlook structural relationships, treating feature vectors as unordered sets. To address this, we propose Graph Matching Optimal Transport (GM-OT), which models linguistic and acoustic sequences as structured graphs. Nodes represent feature embeddings, while edges capture temporal and sequential relationships. GM-OT minimizes both WD (between nodes) and Gromov-Wasserstein distance (GWD) (between edges), leading to a fused Gromov-Wasserstein distance (FGWD) formulation. This enables structured alignment and more efficient knowledge transfer compared to existing OT-based approaches. Theoretical analysis further shows that prior OT-based methods in linguistic knowledge transfer can be viewed as a special case within our GM-OT framework. We evaluate GM-OT on Mandarin ASR using a CTC-based E2E-ASR system with a PLM for knowledge transfer. Experimental results demonstrate significant performance gains over state-of-the-art models, validating the effectiveness of our approach.
Spiking Neural Networks (SNNs) have emerged as a promising approach for energy-efficient and biologically plausible computation. However, due to limitations in existing training methods and inherent model constraints, SNNs often exhibit a performance gap when compared to Artificial Neural Networks (ANNs). Knowledge distillation (KD) has been explored as a technique to transfer knowledge from ANN teacher models to SNN student models to mitigate this gap. Traditional KD methods typically use Kullback-Leibler (KL) divergence to align output distributions. However, conventional KL-based approaches fail to fully exploit the unique characteristics of SNNs, as they tend to overemphasize high-probability predictions while neglecting low-probability ones, leading to suboptimal generalization. To address this, we propose Head-Tail Aware Kullback-Leibler (HTA-KL) divergence, a novel KD method for SNNs. HTA-KL introduces a cumulative probability-based mask to dynamically distinguish between high- and low-probability regions. It assigns adaptive weights to ensure balanced knowledge transfer, enhancing the overall performance. By integrating forward KL (FKL) and reverse KL (RKL) divergence, our method effectively align both head and tail regions of the distribution. We evaluate our methods on CIFAR-10, CIFAR-100 and Tiny ImageNet datasets. Our method outperforms existing methods on most datasets with fewer timesteps.
This work explores the journey towards achieving Bahnaric-Vietnamese translation for the sake of culturally bridging the two ethnic groups in Vietnam. However, translating from Bahnaric to Vietnamese also encounters some difficulties. The most prominent challenge is the lack of available original Bahnaric resources source language, including vocabulary, grammar, dialogue patterns and bilingual corpus, which hinders the data collection process for training. To address this, we leverage a transfer learning approach using sequence-to-sequence pre-training language model. First of all, we leverage a pre-trained Vietnamese language model to capture the characteristics of this language. Especially, to further serve the purpose of machine translation, we aim for a sequence-to-sequence model, not encoder-only like BERT or decoder-only like GPT. Taking advantage of significant similarity between the two languages, we continue training the model with the currently limited bilingual resources of Vietnamese-Bahnaric text to perform the transfer learning from language model to machine translation. Thus, this approach can help to handle the problem of imbalanced resources between two languages, while also optimizing the training and computational processes. Additionally, we also enhanced the datasets using data augmentation to generate additional resources and defined some heuristic methods to help the translation more precise. Our approach has been validated to be highly effective for the Bahnaric-Vietnamese translation model, contributing to the expansion and preservation of languages, and facilitating better mutual understanding between the two ethnic people.
Lifelong learning algorithms enable models to incrementally acquire new knowledge without forgetting previously learned information. Contrarily, the field of machine unlearning focuses on explicitly forgetting certain previous knowledge from pretrained models when requested, in order to comply with data privacy regulations on the right-to-be-forgotten. Enabling efficient lifelong learning with the capability to selectively unlearn sensitive information from models presents a critical and largely unaddressed challenge with contradicting objectives. We address this problem from the perspective of simultaneously preventing catastrophic forgetting and allowing forward knowledge transfer during task-incremental learning, while ensuring exact task unlearning and minimizing memory requirements, based on a single neural network model to be adapted. Our proposed solution, privacy-aware lifelong learning (PALL), involves optimization of task-specific sparse subnetworks with parameter sharing within a single architecture. We additionally utilize an episodic memory rehearsal mechanism to facilitate exact unlearning without performance degradations. We empirically demonstrate the scalability of PALL across various architectures in image classification, and provide a state-of-the-art solution that uniquely integrates lifelong learning and privacy-aware unlearning mechanisms for responsible AI applications.
Analog in-memory computing is a next-generation computing paradigm that promises fast, parallel, and energy-efficient deep learning training and transfer learning (TL). However, achieving this promise has remained elusive due to a lack of suitable training algorithms. Analog memory devices exhibit asymmetric and non-linear switching behavior in addition to device-to-device variation, meaning that most, if not all, of the current off-the-shelf training algorithms cannot achieve good training outcomes. Also, recently introduced algorithms have enjoyed limited attention, as they require bi-directionally switching devices of unrealistically high symmetry and precision and are highly sensitive. A new algorithm chopped TTv2 (c-TTv2), has been introduced, which leverages the chopped technique to address many of the challenges mentioned above. In this paper, we assess the performance of the c-TTv2 algorithm for analog TL using a Swin-ViT model on a subset of the CIFAR100 dataset. We also investigate the robustness of our algorithm to changes in some device specifications, including weight transfer noise, symmetry point skew, and symmetry point variability
Over the past decades, computer-aided diagnosis tools for breast cancer have been developed to enhance screening procedures, yet their clinical adoption remains challenged by data variability and inherent biases. Although foundation models (FMs) have recently demonstrated impressive generalizability and transfer learning capabilities by leveraging vast and diverse datasets, their performance can be undermined by spurious correlations that arise from variations in image quality, labeling uncertainty, and sensitive patient attributes. In this work, we explore the fairness and bias of FMs for breast mammography classification by leveraging a large pool of datasets from diverse sources-including data from underrepresented regions and an in-house dataset. Our extensive experiments show that while modality-specific pre-training of FMs enhances performance, classifiers trained on features from individual datasets fail to generalize across domains. Aggregating datasets improves overall performance, yet does not fully mitigate biases, leading to significant disparities across under-represented subgroups such as extreme breast densities and age groups. Furthermore, while domain-adaptation strategies can reduce these disparities, they often incur a performance trade-off. In contrast, fairness-aware techniques yield more stable and equitable performance across subgroups. These findings underscore the necessity of incorporating rigorous fairness evaluations and mitigation strategies into FM-based models to foster inclusive and generalizable AI.
With the widespread adoption of Large Language Models (LLMs), there is a growing need to establish best practices for leveraging their capabilities beyond traditional natural language tasks. In this paper, a novel cross-domain knowledge transfer framework is proposed to enhance the performance of LLMs in time series forecasting -- a task of increasing relevance in fields such as energy systems, finance, and healthcare. The approach systematically infuses LLMs with structured temporal information to improve their forecasting accuracy. This study evaluates the proposed method on a real-world time series dataset and compares it to a naive baseline where the LLM receives no auxiliary information. Results show that knowledge-informed forecasting significantly outperforms the uninformed baseline in terms of predictive accuracy and generalization. These findings highlight the potential of knowledge transfer strategies to bridge the gap between LLMs and domain-specific forecasting tasks.
We propose a new framework for multi-agent reinforcement learning (MARL), where the agents cooperate in a time-evolving network with latent community structures and mixed memberships. Unlike traditional neighbor-based or fixed interaction graphs, our community-based framework captures flexible and abstract coordination patterns by allowing each agent to belong to multiple overlapping communities. Each community maintains shared policy and value functions, which are aggregated by individual agents according to personalized membership weights. We also design actor-critic algorithms that exploit this structure: agents inherit community-level estimates for policy updates and value learning, enabling structured information sharing without requiring access to other agents' policies. Importantly, our approach supports both transfer learning by adapting to new agents or tasks via membership estimation, and active learning by prioritizing uncertain communities during exploration. Theoretically, we establish convergence guarantees under linear function approximation for both actor and critic updates. To our knowledge, this is the first MARL framework that integrates community structure, transferability, and active learning with provable guarantees.
Air pollution poses a significant threat to public health, causing or exacerbating many respiratory and cardiovascular diseases. In addition, climate change is bringing about more extreme weather events such as wildfires and heatwaves, which can increase levels of pollution and worsen the effects of pollution exposure. Recent advances in personal sensing have transformed the collection of behavioural and physiological data, leading to the potential for new improvements in healthcare. We wish to capitalise on this data, alongside new capabilities in AI for making time series predictions, in order to monitor and predict health outcomes for an individual. Thus, we present a novel workflow for predicting personalised health responses to pollution by integrating physiological data from wearable fitness devices with real-time environmental exposures. The data is collected from various sources in a secure and ethical manner, and is used to train an AI model to predict individual health responses to pollution exposure within a cloud-based, modular framework. We demonstrate that the AI model -- an Adversarial Autoencoder neural network in this case -- accurately reconstructs time-dependent health signals and captures nonlinear responses to pollution. Transfer learning is applied using data from a personal smartwatch, which increases the generalisation abilities of the AI model and illustrates the adaptability of the approach to real-world, user-generated data.
This paper examines two aspects of the isolated sign language recognition (ISLR) task. First, despite the availability of a number of datasets, the amount of data for most individual sign languages is limited. It poses the challenge of cross-language ISLR model training, including transfer learning. Second, similar signs can have different semantic meanings. It leads to ambiguity in dataset labeling and raises the question of the best policy for annotating such signs. To address these issues, this study presents Logos, a novel Russian Sign Language (RSL) dataset, the most extensive ISLR dataset by the number of signers and one of the largest available datasets while also the largest RSL dataset in size and vocabulary. It is shown that a model, pre-trained on the Logos dataset can be used as a universal encoder for other language SLR tasks, including few-shot learning. We explore cross-language transfer learning approaches and find that joint training using multiple classification heads benefits accuracy for the target lowresource datasets the most. The key feature of the Logos dataset is explicitly annotated visually similar sign groups. We show that explicitly labeling visually similar signs improves trained model quality as a visual encoder for downstream tasks. Based on the proposed contributions, we outperform current state-of-the-art results for the WLASL dataset and get competitive results for the AUTSL dataset, with a single stream model processing solely RGB video. The source code, dataset, and pre-trained models are publicly available.
Large-scale pre-trained Vision-Language Models (VLMs) have significantly advanced transfer learning across diverse tasks. However, adapting these models with limited few-shot data often leads to overfitting, undermining their ability to generalize to new tasks. To address this, we propose Multi-Modal Representation Learning (MMRL), which introduces a shared, learnable, modality-agnostic representation space. MMRL generates space tokens projected into both text and image encoders as representation tokens, enabling more effective cross-modal interactions. Unlike prior methods that mainly optimize class token features, MMRL inserts representation tokens into higher encoder layers--where task-specific features are more prominent--while preserving general knowledge in the lower layers. During training, both class and representation features are jointly optimized: a trainable projection layer is applied to representation tokens for task adaptation, while the projection layer for class token remains frozen to retain pre-trained knowledge. To further promote generalization, we introduce a regularization term aligning class and text features with the frozen VLM's zero-shot features. At inference, a decoupling strategy uses both class and representation features for base tasks, but only class features for novel tasks due to their stronger generalization. Building upon this, we propose MMRL++, a parameter-efficient and interaction-aware extension that significantly reduces trainable parameters and enhances intra-modal interactions--particularly across the layers of representation tokens--allowing gradient sharing and instance-specific information to propagate more effectively through the network. Extensive experiments on 15 datasets demonstrate that MMRL and MMRL++ consistently outperform state-of-the-art methods, achieving a strong balance between task-specific adaptation and generalization.