Large Language Models (LLMs) are capable of reasoning over diverse input data modalities through pre-trained encoders. However, the growing diversity of input data modalities prevents incorporating all modalities into LLMs, especially when LLMs are deployed on resource-constrained edge devices for embodied AI applications. Instead, a better option is to adaptively involve only the useful modalities at runtime, depending on the current environmental contexts and task requirements. For such modality adaptation, existing work adopts fixed connections between encoders and the LLM's input layer, leading to high training cost at runtime and ineffective cross-modal interaction. In this paper, we address these limitations by presenting mPnP-LLM, a new technique that allows fully elastic, automated and prompt runtime modality adaptation, by connecting unimodal encoders to a flexible set of last LLM blocks and making such latent connections fully trainable at runtime. Experiments over the nuScenes-QA dataset show that mPnP-LLM can achieve up to 3.7x FLOPs reduction and 30% GPU memory usage reduction, while retaining on-par accuracy with the existing schemes. Under the same compute budget, mPnP-LLM improves the task accuracy by up to 4% compared to the best existing scheme.
In this work, we use multi-view aerial images to reconstruct the geometry, lighting, and material of facades using neural signed distance fields (SDFs). Without the requirement of complex equipment, our method only takes simple RGB images captured by a drone as inputs to enable physically based and photorealistic novel-view rendering, relighting, and editing. However, a real-world facade usually has complex appearances ranging from diffuse rocks with subtle details to large-area glass windows with specular reflections, making it hard to attend to everything. As a result, previous methods can preserve the geometry details but fail to reconstruct smooth glass windows or verse vise. In order to address this challenge, we introduce three spatial- and semantic-adaptive optimization strategies, including a semantic regularization approach based on zero-shot segmentation techniques to improve material consistency, a frequency-aware geometry regularization to balance surface smoothness and details in different surfaces, and a visibility probe-based scheme to enable efficient modeling of the local lighting in large-scale outdoor environments. In addition, we capture a real-world facade aerial 3D scanning image set and corresponding point clouds for training and benchmarking. The experiment demonstrates the superior quality of our method on facade holistic inverse rendering, novel view synthesis, and scene editing compared to state-of-the-art baselines.
Robotic ophthalmic surgery is an emerging technology to facilitate high-precision interventions such as retina penetration in subretinal injection and removal of floating tissues in retinal detachment depending on the input imaging modalities such as microscopy and intraoperative OCT (iOCT). Although iOCT is explored to locate the needle tip within its range-limited ROI, it is still difficult to coordinate iOCT's motion with the needle, especially at the initial target-approaching stage. Meanwhile, due to 2D perspective projection and thus the loss of depth information, current image-based methods cannot effectively estimate the needle tip's trajectory towards both retinal and floating targets. To address this limitation, we propose to use the shadow positions of the target and the instrument tip to estimate their relative depth position and accordingly optimize the instrument tip's insertion trajectory until the tip approaches targets within iOCT's scanning area. Our method succeeds target approaching on a retina model, and achieves an average depth error of 0.0127 mm and 0.3473 mm for floating and retinal targets respectively in the surgical simulator without damaging the retina.
In this paper, we investigate the problem of fast spectrum sharing in vehicle-to-everything communication. In order to improve the spectrum efficiency of the whole system, the spectrum of vehicle-to-infrastructure links is reused by vehicle-to-vehicle links. To this end, we model it as a problem of deep reinforcement learning and tackle it with proximal policy optimization. A considerable number of interactions are often required for training an agent with good performance, so simulation-based training is commonly used in communication networks. Nevertheless, severe performance degradation may occur when the agent is directly deployed in the real world, even though it can perform well on the simulator, due to the reality gap between the simulation and the real environments. To address this issue, we make preliminary efforts by proposing an algorithm based on meta reinforcement learning. This algorithm enables the agent to rapidly adapt to a new task with the knowledge extracted from similar tasks, leading to fewer interactions and less training time. Numerical results show that our method achieves near-optimal performance and exhibits rapid convergence.
Fine-tuning is the most effective way of adapting pre-trained large language models (LLMs) to downstream applications. With the fast growth of LLM-enabled AI applications and democratization of open-souced LLMs, fine-tuning has become possible for non-expert individuals, but intensively performed LLM fine-tuning worldwide could result in significantly high energy consumption and carbon footprint, which may bring large environmental impact. Mitigating such environmental impact towards Green AI directly correlates to reducing the FLOPs of fine-tuning, but existing techniques on efficient LLM fine-tuning can only achieve limited reduction of such FLOPs, due to their ignorance of the backpropagation cost in fine-tuning. To address this limitation, in this paper we present GreenTrainer, a new LLM fine-tuning technique that adaptively evaluates different tensors' backpropagation costs and contributions to the fine-tuned model accuracy, to minimize the fine-tuning cost by selecting the most appropriate set of tensors in training. Such selection in GreenTrainer is made based on a given objective of FLOPs reduction, which can flexibly adapt to the carbon footprint in energy supply and the need in Green AI. Experiment results over multiple open-sourced LLM models and abstractive summarization datasets show that, compared to fine-tuning the whole LLM model, GreenTrainer can save up to 64% FLOPs in fine-tuning without any noticeable model accuracy loss. Compared to the existing fine-tuning techniques such as LoRa, GreenTrainer can achieve up to 4% improvement on model accuracy with on-par FLOPs reduction.
Generalized Few-shot Semantic Segmentation (GFSS) extends Few-shot Semantic Segmentation (FSS) to simultaneously segment unseen classes and seen classes during evaluation. Previous works leverage additional branch or prototypical aggregation to eliminate the constrained setting of FSS. However, representation division and embedding prejudice, which heavily results in poor performance of GFSS, have not been synthetical considered. We address the aforementioned problems by jointing the prototypical kernel learning and open-set foreground perception. Specifically, a group of learnable kernels is proposed to perform segmentation with each kernel in charge of a stuff class. Then, we explore to merge the prototypical learning to the update of base-class kernels, which is consistent with the prototype knowledge aggregation of few-shot novel classes. In addition, a foreground contextual perception module cooperating with conditional bias based inference is adopted to perform class-agnostic as well as open-set foreground detection, thus to mitigate the embedding prejudice and prevent novel targets from being misclassified as background. Moreover, we also adjust our method to the Class Incremental Few-shot Semantic Segmentation (CIFSS) which takes the knowledge of novel classes in a incremental stream. Extensive experiments on PASCAL-5i and COCO-20i datasets demonstrate that our method performs better than previous state-of-the-art.
Local geometric information, i.e. normal and point distribution, is crucial for LiDAR-based simultaneous localization and mapping (SLAM) because it provides constrains for data association, which further determines the direction of optimization and ultimately affects the accuracy of poses. However, estimating normal and point distribution are time-consuming tasks even with the assistance of the KDtree or volumetic maps. To achieve fast normal estimation, we look into the structural information of LiDAR scan and propose a novel fast approximate least squares (FALS) method. With the pre-computed bearing information, estimating the normal requires only the range information of the points when a new scan arrives. To efficiently estimate the distribution of points, we extend the ikd-tree to manage the map in voxels and update its point cloud distribution incrementally while maintaining its consistency with the normals. For scan points that satisfy visibility and consistency checks based on normal, we devise a robust and accurate hierarchical data association schema considering the distribution where point-to-surfel is prioritized over point-to-plane. We further fix voxels after the distribution convergences to balance the time consumption and the correctness of representation. Extensive experiments on diverse public datasets demonstrate the advantages of our system compared to other state-of-the-art methods.
The growing interest in language-conditioned robot manipulation aims to develop robots capable of understanding and executing complex tasks, with the objective of enabling robots to interpret language commands and manipulate objects accordingly. While language-conditioned approaches demonstrate impressive capabilities for addressing tasks in familiar environments, they encounter limitations in adapting to unfamiliar environment settings. In this study, we propose a general-purpose, language-conditioned approach that combines base skill priors and imitation learning under unstructured data to enhance the algorithm's generalization in adapting to unfamiliar environments. We assess our model's performance in both simulated and real-world environments using a zero-shot setting. In the simulated environment, the proposed approach surpasses previously reported scores for CALVIN benchmark, especially in the challenging Zero-Shot Multi-Environment setting. The average completed task length, indicating the average number of tasks the agent can continuously complete, improves more than 2.5 times compared to the state-of-the-art method HULC. In addition, we conduct a zero-shot evaluation of our policy in a real-world setting, following training exclusively in simulated environments without additional specific adaptations. In this evaluation, we set up ten tasks and achieved an average 30% improvement in our approach compared to the current state-of-the-art approach, demonstrating a high generalization capability in both simulated environments and the real world. For further details, including access to our code and videos, please refer to https://demoviewsite.wixsite.com/spil
Generative model-based deep clustering frameworks excel in classifying complex data, but are limited in handling dynamic and complex features because they require prior knowledge of the number of clusters. In this paper, we propose a nonparametric deep clustering framework that employs an infinite mixture of Gaussians as a prior. Our framework utilizes a memoized online variational inference method that enables the "birth" and "merge" moves of clusters, allowing our framework to cluster data in a "dynamic-adaptive" manner, without requiring prior knowledge of the number of features. We name the framework as DIVA, a Dirichlet Process-based Incremental deep clustering framework via Variational Auto-Encoder. Our framework, which outperforms state-of-the-art baselines, exhibits superior performance in classifying complex data with dynamically changing features, particularly in the case of incremental features.