Pre-trained Visual-Language Models (VLMs) have demonstrated powerful performance on various downstream tasks. Recently, many prompt tuning methods represented by Context Optimization (CoOp) have effectively adapted VLMs to few-shot tasks. However, the CoOp-based methods suffer from overfitting to base classes, which impairs the model’s generalization to new classes. Considering that meta-learning excels at generalizing to new classes, we combine meta-learning with CoOp-like vision-language model fine-tuning methods to improve performance on few-shot generation tasks. In this paper, we present a novel Meta-learning-based Multi-Textual Prompt tuning (MMTP) method, which learns multiple textual prompts leveraging meta-learning to enhance the visual-language model’s representation and generalization capabilities. Specifically, we introduce multi-textual prompts to enhance the representation of the model for improving the recognition of base classes. Simultaneously, we employ meta-learning to optimize prompt training, bolstering the model’s generalization to new classes. Extensive experiments demonstrate the superiority of our method under base-to-new generalization and cross-domain generalization settings. Furthermore, we also conduct ablation studies to validate the effectiveness of each component.

Low-light vision remains a fundamental challenge in computer vision due to severe illumination degradation, which significantly affects the performance of downstream tasks such as detection and segmentation. While recent state-of-the-art methods have improved performance through invariant feature learning modules, they still fall short due to incomplete modeling of low-light conditions. Therefore, we revisit low-light image formation and extend the classical Lambertian model to better characterize low-light conditions. By shifting our analysis to the frequency domain, we theoretically prove that the frequency-domain channel ratio can be leveraged to extract illumination-invariant features via a structured filtering process. We then propose a novel and end-to-end trainable module named \textbf{F}requency-domain \textbf{R}adial \textbf{B}asis \textbf{Net}work (\textbf{FRBNet}), which integrates the frequency-domain channel ratio operation with a learnable frequency domain filter for the overall illumination-invariant feature enhancement. As a plug-and-play module, FRBNet can be integrated into existing networks for low-light downstream tasks without modifying loss functions. Extensive experiments across various downstream tasks demonstrate that FRBNet achieves superior performance, including +2.2 mAP for dark object detection and +2.9 mIoU for nighttime segmentation. Code is available at: \url{https://github.com/Sing-Forevet/FRBNet}.

Model-based offline reinforcement learning (RL) has made remarkable progress, offering a promising avenue for improving generalization with synthetic model rollouts. Existing works primarily focus on incorporating pessimism for policy optimization, usually via constructing a Pessimistic Markov Decision Process (P-MDP). However, the P-MDP discourages the policies from learning in out-of-distribution (OOD) regions beyond the support of offline datasets, which can under-utilize the generalization ability of dynamics models. In contrast, we propose constructing an Optimistic MDP (O-MDP). We initially observed the potential benefits of optimism brought by encouraging more OOD rollouts. Motivated by this observation, we present ORPO, a simple yet effective model-based offline RL framework. ORPO generates Optimistic model Rollouts for Pessimistic offline policy Optimization. Specifically, we train an optimistic rollout policy in the O-MDP to sample more OOD model rollouts. Then we relabel the sampled state-action pairs with penalized rewards, and optimize the output policy in the P-MDP. Theoretically, we demonstrate that the performance of policies trained with ORPO can be lower-bounded in linear MDPs. Experimental results show that our framework significantly outperforms P-MDP baselines by a margin of 30%, achieving state-of-the-art performance on the widely-used benchmark. Moreover, ORPO exhibits notable advantages in problems that require generalization.

Pre-trained large language models exhibit significant potential in speech and language processing. Fine-tuning all parameters becomes impractical when confronted with numerous downstream tasks. To address this challenge, various low-rank adaptation techniques have been introduced for parameter-efficient fine-tuning, which freeze the over-parametrized models and learn incremental parameter updates within smaller subspaces. However, our observation reveals that most directions of the learned subspace play a minor role in the incremental updates. Consequently, fine-tuned models may not achieve optimal performance. To bridge this gap, we introduce NNM-LoRA, which strives to harness more meaningful singular directions. Through Nuclear Norm Maximization (NNM), we can better regulate the allocation of singular values. Accordingly, we propose a parameter-free plug-and-play regularizer for low-rank updates. This innovative approach allows us to utilize as many singular directions of the subspace as possible during the training of low-rank updates. To validate the effectiveness of NNM-LoRA, we conduct extensive experiments involving different pre-trained models on various natural language understanding tasks. Results demonstrate that NNM-LoRA exhibits significant improvements compared to baseline methods.

Vision Transformer (ViT) models have recently emerged as powerful and versatile tools for various visual tasks. In this article, we investigate ViT in a more challenging scenario within the context of few-shot conditions. Recent work has achieved promising results in few-shot image classification by utilizing pre-trained vision transformer models. However, this work employs full fine-tuning for the downstream tasks, leading to significant overfitting and storage issues, especially in the remote sensing domain. In order to tackle these issues, we turn to the recently proposed Parameter-Efficient Tuning (PETuning) methods, which update only the newly added parameters while keeping the pre-trained backbone frozen. Inspired by these methods, we propose the Meta Visual Prompt Tuning (MVP) method. Specifically, we integrate the prompt-tuning-based PETuning method into the meta-learning framework and tailor it for remote sensing datasets, resulting in an efficient framework for Few-Shot Remote Sensing Scene Classification (FS-RSSC). Moreover, we introduce a novel data augmentation scheme that exploits patch embedding recombination to enhance the data diversity and quantity. This scheme is generalizable to any network that employs the ViT architecture as its backbone. Experimental results on the FS-RSSC benchmark demonstrate the superior performance of the proposed MVP over existing methods in various settings, including various-way-various-shot, various-way-one-shot, and cross-domain adaptation.

Federated learning (FL) enables distributed participants to collectively learn a strong global model without sacrificing their individual data privacy. Mainstream FL approaches require each participant to share a common network architecture and further assume that data are sampled IID across participants. However, in real-world deployments, participants may require heterogeneous network architectures; and the data distribution is almost non-uniform. To address these issues we introduce FedH2L, which is agnostic to the model architecture and robust to different data distributions across participants. In contrast to approaches sharing parameters or gradients, FedH2L relies on mutual distillation, exchanging only posteriors on a shared seed set between participants in a decentralized manner. This makes it extremely bandwidth efficient, model agnostic, and crucially produces models capable of performing well on the whole data distribution when learning from heterogeneous silos.

Off-Policy Actor-Critic (OffP-AC) methods have proven successful in a variety of continuous control tasks. Normally, the critic's action-value function is updated using temporal-difference, and the critic in turn provides a loss for the actor that trains it to take actions with higher expected return. In this paper, we introduce a flexible and augmented meta-critic that observes the learning process and meta-learns an additional loss for the actor that accelerates and improves actor-critic learning. Compared to existing meta-learning algorithms, meta-critic is rapidly learned online for a single task, rather than slowly over a family of tasks. Crucially, our meta-critic is designed for off-policy based learners, which currently provide state-of-the-art reinforcement learning sample efficiency. We demonstrate that online meta-critic learning benefits to a variety of continuous control tasks when combined with contemporary OffP-AC methods DDPG, TD3 and SAC.

The well known domain shift issue causes model performance to degrade when deployed to a new target domain with different statistics to training. Domain adaptation techniques alleviate this, but need some instances from the target domain to drive adaptation. Domain generalisation is the recently topical problem of learning a model that generalises to unseen domains out of the box, and various approaches aim to train a domain-invariant feature extractor, typically by adding some manually designed losses. In this work, we propose a learning to learn approach, where the auxiliary loss that helps generalisation is itself learned. Beyond conventional domain generalisation, we consider a more challenging setting of heterogeneous domain generalisation, where the unseen domains do not share label space with the seen ones, and the goal is to train a feature representation that is useful off-the-shelf for novel data and novel categories. Experimental evaluation demonstrates that our method outperforms state-of-the-art solutions in both settings.

Following trails in the wild is an essential capability of out-door autonomous mobile robots. Recently, deep learningbased approaches have made great advancements in this field. However, the existing research only focuses on the trail following with a single robot. In contrast, many robotic tasks in the reality, such as search and patrolling, are conducted by a group of robots. While these robots are grouped to move in the wild, they can cooperate to significantly promote the trail following accuracy, for example, by sharing images of different view angles or real-time decision fusion. This paper proposes such an approach named DL-Cooper that enables multi-robot visionbased trail following based on deep learning algorithms. It allows each robot to make a decision respectively with deep neural network and then fusion the decisions on the collective level with the support of back-end cloud computing infrastructure. It also takes Quality of Service (QoS) assurance, a very essential property of robotic software, into consideration. By limiting the condition to fusion decisions, the time latency can be minimally sacrificed. Experiments on the real-world dataset show that our approach has significantly improved the accuracy of the singlerobot system.

Uncertainty is a great challenge for environment perception of autonomous robots. For instance, while building semantic maps (i.e., maps with semantic labels such as object names), the robot may encounter unexpected objects of which it has no knowledge. It will lead to inevitable failures in traditional environment modeling software. The abundant knowledge being accumulated on the Internet has the potential to assist robots to handle such kind of uncertainly. However, existing researches have not touched this issue yet. This paper proposes a cloud-based semantic mapping engine named SemaCloud, which can not only augment robot's environment modeling capability by the rich cloud resources but also cope with uncertainty by utilizing the Internet knowledge on necessary. It adopts a state-of-art Deep Neural Network (DNN) for real-time and accurate recognition of pre-trained objects. If an object is beyond the knowledge of this DNN, a special mechanism named QoS-aware cloud phase transition is triggered to seek help from existing recognition services on the Internet. By a set of carefully-designed algorithms, it can maximize benefits and minimize the negative impacts on the Quality of Service (QoS) properties of robotic applications, which is essential to many robot scenarios. The experiments on both open datasets and real robots show that our work can handle uncertainly successfully in robotic semantic mapping without sacrificing critical real-time constraints.