The quadratic complexity and weak length extrapolation of Transformers limits their ability to scale to long sequences, and while sub-quadratic solutions like linear attention and state space models exist, they empirically underperform Transformers in pretraining efficiency and downstream task accuracy. We introduce Megalodon, a neural architecture for efficient sequence modeling with unlimited context length. Megalodon inherits the architecture of Mega (exponential moving average with gated attention), and further introduces multiple technical components to improve its capability and stability, including complex exponential moving average (CEMA), timestep normalization layer, normalized attention mechanism and pre-norm with two-hop residual configuration. In a controlled head-to-head comparison with Llama2, Megalodon achieves better efficiency than Transformer in the scale of 7 billion parameters and 2 trillion training tokens. Megalodon reaches a training loss of 1.70, landing mid-way between Llama2-7B (1.75) and 13B (1.67). Code: https://github.com/XuezheMax/megalodon
Mixture-of-Expert (MoE) based large language models (LLMs), such as the recent Mixtral and DeepSeek-MoE, have shown great promise in scaling model size without suffering from the quadratic growth of training cost of dense transformers. Like dense models, training MoEs requires answering the same question: given a training budget, what is the optimal allocation on the model size and number of tokens? We study the scaling law of MoE-based LLMs regarding the relations between the model performance, model size, dataset size, and the expert degree. Echoing previous research studying MoE in different contexts, we observe the diminishing return of increasing the number of experts, but this seems to suggest we should scale the number of experts until saturation, as the training cost would remain constant, which is problematic during inference time. We propose to amend the scaling law of MoE by introducing inference efficiency as another metric besides the validation loss. We find that MoEs with a few (4/8) experts are the most serving efficient solution under the same performance, but costs 2.5-3.5x more in training. On the other hand, training a (16/32) expert MoE much smaller (70-85%) than the loss-optimal solution, but with a larger training dataset is a promising setup under a training budget.
Wireless signal recognition (WSR) is crucial in modern and future wireless communication networks since it aims to identify properties of the received signal. Although many deep learning-based WSR models have been developed, they still rely on a large amount of labeled training data. Thus, they cannot tackle the few-sample problem in the practically and dynamically changing wireless communication environment. To overcome this challenge, a novel SSwsrNet framework is proposed by using the deep residual shrinkage network (DRSN) and semi-supervised learning. The DRSN can learn discriminative features from noisy signals. Moreover, a modular semi-supervised learning method that combines labeled and unlabeled data using MixMatch is exploited to further improve the classification performance under few-sample conditions. Extensive simulation results on automatic modulation classification (AMC) and wireless technology classification (WTC) demonstrate that our proposed WSR scheme can achieve better performance than the benchmark schemes in terms of classification accuracy. This novel method enables more robust and adaptive signal recognition for next-generation wireless networks.
We present a differentiable rendering framework to learn structured 3D abstractions in the form of primitive assemblies from sparse RGB images capturing a 3D object. By leveraging differentiable volume rendering, our method does not require 3D supervision. Architecturally, our network follows the general pipeline of an image-conditioned neural radiance field (NeRF) exemplified by pixelNeRF for color prediction. As our core contribution, we introduce differential primitive assembly (DPA) into NeRF to output a 3D occupancy field in place of density prediction, where the predicted occupancies serve as opacity values for volume rendering. Our network, coined DPA-Net, produces a union of convexes, each as an intersection of convex quadric primitives, to approximate the target 3D object, subject to an abstraction loss and a masking loss, both defined in the image space upon volume rendering. With test-time adaptation and additional sampling and loss designs aimed at improving the accuracy and compactness of the obtained assemblies, our method demonstrates superior performance over state-of-the-art alternatives for 3D primitive abstraction from sparse views.
Multi-robot collaborative navigation is an essential ability where teamwork and synchronization are keys. In complex and uncertain environments, adaptive formation is vital, as rigid formations prove to be inadequate. The ability of robots to dynamically adjust their formation enables navigation through unpredictable spaces, maintaining cohesion, and effectively responding to environmental challenges. In this paper, we introduce a novel approach that uses bi-level learning framework. Specifically, we use graph learning at a high level for group coordination and reinforcement learning for individual navigation. We innovate by integrating a spring-damper model within the reinforcement learning reward mechanism, addressing the rigidity of traditional formation control methods. During execution, our approach enables a team of robots to successfully navigate challenging environments, maintain a desired formation shape, and dynamically adjust their formation scale based on environmental information. We conduct extensive experiments to evaluate our approach across three distinct formation scenarios in multi-robot navigation: circle, line, and wedge. Experimental results show that our approach achieves promising results and scalability on multi-robot navigation with formation adaptation.
Previous multi-task dense prediction methods based on the Mixture of Experts (MoE) have received great performance but they neglect the importance of explicitly modeling the global relations among all tasks. In this paper, we present a novel decoder-focused method for multi-task dense prediction, called Mixture-of-Low-Rank-Experts (MLoRE). To model the global task relationships, MLoRE adds a generic convolution path to the original MoE structure, where each task feature can go through this path for explicit parameter sharing. Furthermore, to control the parameters and computational cost brought by the increase in the number of experts, we take inspiration from LoRA and propose to leverage the low-rank format of a vanilla convolution in the expert network. Since the low-rank experts have fewer parameters and can be dynamically parameterized into the generic convolution, the parameters and computational cost do not change much with the increase of experts. Benefiting from this design, we increase the number of experts and its reception field to enlarge the representation capacity, facilitating multiple dense tasks learning in a unified network. Extensive experiments on the PASCAL-Context and NYUD-v2 benchmarks show that our MLoRE achieves superior performance compared to previous state-of-the-art methods on all metrics. Our code is available at https://github.com/YuqiYang213/MLoRE.
Image fusion aims to combine information from different source images to create a comprehensively representative image. Existing fusion methods are typically helpless in dealing with degradations in low-quality source images and non-interactive to multiple subjective and objective needs. To solve them, we introduce a novel approach that leverages semantic text guidance image fusion model for degradation-aware and interactive image fusion task, termed as Text-IF. It innovatively extends the classical image fusion to the text guided image fusion along with the ability to harmoniously address the degradation and interaction issues during fusion. Through the text semantic encoder and semantic interaction fusion decoder, Text-IF is accessible to the all-in-one infrared and visible image degradation-aware processing and the interactive flexible fusion outcomes. In this way, Text-IF achieves not only multi-modal image fusion, but also multi-modal information fusion. Extensive experiments prove that our proposed text guided image fusion strategy has obvious advantages over SOTA methods in the image fusion performance and degradation treatment. The code is available at https://github.com/XunpengYi/Text-IF.
Multi-modal large language models (MLLMs) can understand image-language prompts and demonstrate impressive reasoning ability. In this paper, we extend MLLMs' output by empowering MLLMs with the segmentation ability. The extended MLLMs can both output language responses to the image-language prompts and segment the regions that the complex question or query in the language prompts focuses on. To this end, the existing work, LISA, enlarges the original word embeddings with an additional segment token and fine-tunes dialogue generation and query-focused segmentation together, where the feature of the segment token is used to prompt the segment-anything model. Although they achieve superior segmentation performance, we observe that the dialogue ability decreases by a large margin compared to the original MLLMs. To maintain the original MLLMs' dialogue ability, we propose a novel MLLMs framework, coined as LLaVASeg, which leverages a chain-of-thought prompting strategy to instruct the MLLMs to segment the target region queried by the user. The MLLMs are first prompted to reason about the simple description of the target region from the complicated user query, then extract the visual attributes of the target region according to the understanding of MLLMs to the image. These visual attributes, such as color and relative locations, are utilized to prompt the downstream segmentation model. Experiments show that the proposed method keeps the original dialogue ability and equips the MLLMs' model with strong reasoning segmentation ability. The code is available at https://github.com/YuqiYang213/LLaVASeg.
In this paper, we propose a simple and strong framework for Tracking Any Point with TRansformers (TAPTR). Based on the observation that point tracking bears a great resemblance to object detection and tracking, we borrow designs from DETR-like algorithms to address the task of TAP. In the proposed framework, in each video frame, each tracking point is represented as a point query, which consists of a positional part and a content part. As in DETR, each query (its position and content feature) is naturally updated layer by layer. Its visibility is predicted by its updated content feature. Queries belonging to the same tracking point can exchange information through self-attention along the temporal dimension. As all such operations are well-designed in DETR-like algorithms, the model is conceptually very simple. We also adopt some useful designs such as cost volume from optical flow models and develop simple designs to provide long temporal information while mitigating the feature drifting issue. Our framework demonstrates strong performance with state-of-the-art performance on various TAP datasets with faster inference speed.
Advancements in self-supervised pre-training (SSL) have significantly advanced the field of learning transferable time series representations, which can be very useful in enhancing the downstream task. Despite being effective, most existing works struggle to achieve cross-domain SSL pre-training, missing valuable opportunities to integrate patterns and features from different domains. The main challenge lies in the significant differences in the characteristics of time-series data across different domains, such as variations in the number of channels and temporal resolution scales. To address this challenge, we propose CrossTimeNet, a novel cross-domain SSL learning framework to learn transferable knowledge from various domains to largely benefit the target downstream task. One of the key characteristics of CrossTimeNet is the newly designed time series tokenization module, which could effectively convert the raw time series into a sequence of discrete tokens based on a reconstruction optimization process. Besides, we highlight that predicting a high proportion of corrupted tokens can be very helpful for extracting informative patterns across different domains during SSL pre-training, which has been largely overlooked in past years. Furthermore, unlike previous works, our work treats the pre-training language model (PLM) as the initialization of the encoder network, investigating the feasibility of transferring the knowledge learned by the PLM to the time series area. Through these efforts, the path to cross-domain pre-training of a generic time series model can be effectively paved. We conduct extensive experiments in a real-world scenario across various time series classification domains. The experimental results clearly confirm CrossTimeNet's superior performance.