Language models demonstrate remarkable capacity to generalize representations learned in one modality to downstream tasks in other modalities. Can we trace this ability to individual neurons? We study the case where a frozen text transformer is augmented with vision using a self-supervised visual encoder and a single linear projection learned on an image-to-text task. Outputs of the projection layer are not immediately decodable into language describing image content; instead, we find that translation between modalities occurs deeper within the transformer. We introduce a procedure for identifying "multimodal neurons" that convert visual representations into corresponding text, and decoding the concepts they inject into the model's residual stream. In a series of experiments, we show that multimodal neurons operate on specific visual concepts across inputs, and have a systematic causal effect on image captioning.
相關內容
The recently introduced graphical continuous Lyapunov models provide a new approach to statistical modeling of correlated multivariate data. The models view each observation as a one-time cross-sectional snapshot of a multivariate dynamic process in equilibrium. The covariance matrix for the data is obtained by solving a continuous Lyapunov equation that is parametrized by the drift matrix of the dynamic process. In this context, different statistical models postulate different sparsity patterns in the drift matrix, and it becomes a crucial problem to clarify whether a given sparsity assumption allows one to uniquely recover the drift matrix parameters from the covariance matrix of the data. We study this identifiability problem by representing sparsity patterns by directed graphs. Our main result proves that the drift matrix is globally identifiable if and only if the graph for the sparsity pattern is simple (i.e., does not contain directed two-cycles). Moreover, we present a necessary condition for generic identifiability and provide a computational classification of small graphs with up to 5 nodes.
Recent endeavors in video temporal grounding enforce strong cross-modal interactions through attention mechanisms to overcome the modality gap between video and text query. However, previous works treat all video clips equally regardless of their semantic relevance with the text query in attention modules. In this paper, our goal is to provide clues for query-associated video clips within the crossmodal encoding process. With our Correlation-Guided Detection Transformer~(CG-DETR), we explore the appropriate clip-wise degree of cross-modal interactions and how to exploit such degrees for prediction. First, we design an adaptive cross-attention layer with dummy tokens. Dummy tokens conditioned by text query take a portion of the attention weights, preventing irrelevant video clips from being represented by the text query. Yet, not all word tokens equally inherit the text query's correlation to video clips. Thus, we further guide the cross-attention map by inferring the fine-grained correlation between video clips and words. We enable this by learning a joint embedding space for high-level concepts, i.e., moment and sentence level, and inferring the clip-word correlation. Lastly, we use a moment-adaptive saliency detector to exploit each video clip's degrees of text engagement. We validate the superiority of CG-DETR with the state-of-the-art results on various benchmarks for both moment retrieval and highlight detection. Codes are available at //github.com/wjun0830/CGDETR.
Recovering masked feedback with neural models is a popular paradigm in recommender systems. Seeing the success of diffusion models in solving ill-posed inverse problems, we introduce a conditional diffusion framework for collaborative filtering that iteratively reconstructs a user's hidden preferences guided by its historical interactions. To better align with the intrinsic characteristics of implicit feedback data, we implement forward diffusion by applying synthetic smoothing filters to interaction signals on an item-item graph. The resulting reverse diffusion can be interpreted as a personalized process that gradually refines preference scores. Through graph Fourier transform, we equivalently characterize this model as an anisotropic Gaussian diffusion in the graph spectral domain, establishing both forward and reverse formulations. Our model outperforms state-of-the-art methods by a large margin on one dataset and yields competitive results on the others.
Probabilistic couplings are the foundation for many probabilistic relational program logics and arise when relating random sampling statements across two programs. In relational program logics, this manifests as dedicated coupling rules that, e.g., say we may reason as if two sampling statements return the same value. However, this approach fundamentally requires aligning or "synchronizing" the sampling statements of the two programs which is not always possible. In this paper, we develop Clutch, a higher-order probabilistic relational separation logic that addresses this issue by supporting asynchronous probabilistic couplings. We use Clutch to develop a logical step-indexed logical relational to reason about contextual refinement and equivalence of higher-order programs written in a rich language with higher-order local state and impredicative polymorphism. Finally, we demonstrate the usefulness of our approach on a number of case studies. All the results that appear in the paper have been formalized in the Coq proof assistant using the Coquelicot library and the Iris separation logic framework.
Frequency-based methods have been successfully employed in creating high fidelity data-driven reduced order models (DDROMs) for linear dynamical systems. These methods require access to values (and sometimes derivatives) of the frequency-response function (transfer function) in the complex plane. These frequency domain values can at times be costly or difficult to obtain (especially if the method of choice requires resampling); instead one may have access to only time-domain input-output data. The data informativity approach to moment matching provides a powerful new framework for recovering the required frequency data from a single time-domain trajectory. In this work, we analyze and extend upon this framework, resulting in vastly improved conditioning of the associated linear systems, an error indicator, and removal of an assumption that the system order is known. This analysis leads to a robust algorithm for recovering frequency information from time-domain data, suitable for large scale systems. We demonstrate the effectiveness of our algorithm by forming frequency based DDROMs from time-domain data of several dynamical systems.
The human ability to learn, generalize, and control complex manipulation tasks through multi-modality feedback suggests a unique capability, which we refer to as dexterity intelligence. Understanding and assessing this intelligence is a complex task. Amidst the swift progress and extensive proliferation of large language models (LLMs), their applications in the field of robotics have garnered increasing attention. LLMs possess the ability to process and generate natural language, facilitating efficient interaction and collaboration with robots. Researchers and engineers in the field of robotics have recognized the immense potential of LLMs in enhancing robot intelligence, human-robot interaction, and autonomy. Therefore, this comprehensive review aims to summarize the applications of LLMs in robotics, delving into their impact and contributions to key areas such as robot control, perception, decision-making, and path planning. We first provide an overview of the background and development of LLMs for robotics, followed by a description of the benefits of LLMs for robotics and recent advancements in robotics models based on LLMs. We then delve into the various techniques used in the model, including those employed in perception, decision-making, control, and interaction. Finally, we explore the applications of LLMs in robotics and some potential challenges they may face in the near future. Embodied intelligence is the future of intelligent science, and LLMs-based robotics is one of the promising but challenging paths to achieve this.
This letter investigates the secure delivery performance of the rate-splitting multiple access scheme in land mobile satellite (LMS) systems, considering that the private messages intended by a terminal can be eavesdropped by any others from the broadcast signals. Specifically, the considered system has an N-antenna satellite and numerous single-antenna land users. Maximum ratio transmission (MRT) and matched-filtering (MF) precoding techniques are adopted at the satellite separately for the common messages (CMs) and for the private messages (PMs), which are both implemented based on the estimated LMS channels suffering from the Shadowed-Rician fading. Then, closed-form expressions are derived for the ergodic rates for decoding the CM, and for decoding the PM at the intended user respectively, and more importantly, we also derive the ergodic secrecy rate against eavesdropping. Finally, numerical results are provided to validate the correctness of the proposed analysis models, as well as to show some interesting comparisons.
Convolution models with long filters have demonstrated state-of-the-art reasoning abilities in many long-sequence tasks but lag behind the most optimized Transformers in wall-clock time. A major bottleneck is the Fast Fourier Transform (FFT)--which allows long convolutions to run in $O(N logN)$ time in sequence length $N$ but has poor hardware utilization. In this paper, we study how to optimize the FFT convolution. We find two key bottlenecks: the FFT does not effectively use specialized matrix multiply units, and it incurs expensive I/O between layers of the memory hierarchy. In response, we propose FlashFFTConv. FlashFFTConv uses a matrix decomposition that computes the FFT using matrix multiply units and enables kernel fusion for long sequences, reducing I/O. We also present two sparse convolution algorithms--1) partial convolutions and 2) frequency-sparse convolutions--which can be implemented simply by skipping blocks in the matrix decomposition, enabling further opportunities for memory and compute savings. FlashFFTConv speeds up exact FFT convolutions by up to 7.93$\times$ over PyTorch and achieves up to 4.4$\times$ speedup end-to-end. Given the same compute budget, FlashFFTConv allows Hyena-GPT-s to achieve 2.3 points better perplexity on the PILE and M2-BERT-base to achieve 3.3 points higher GLUE score--matching models with twice the parameter count. FlashFFTConv also achieves 96.1% accuracy on Path-512, a high-resolution vision task where no model had previously achieved better than 50%. Furthermore, partial convolutions enable longer-sequence models--yielding the first DNA model that can process the longest human genes (2.3M base pairs)--and frequency-sparse convolutions speed up pretrained models while maintaining or improving model quality.
Denoising diffusion models represent a recent emerging topic in computer vision, demonstrating remarkable results in the area of generative modeling. A diffusion model is a deep generative model that is based on two stages, a forward diffusion stage and a reverse diffusion stage. In the forward diffusion stage, the input data is gradually perturbed over several steps by adding Gaussian noise. In the reverse stage, a model is tasked at recovering the original input data by learning to gradually reverse the diffusion process, step by step. Diffusion models are widely appreciated for the quality and diversity of the generated samples, despite their known computational burdens, i.e. low speeds due to the high number of steps involved during sampling. In this survey, we provide a comprehensive review of articles on denoising diffusion models applied in vision, comprising both theoretical and practical contributions in the field. First, we identify and present three generic diffusion modeling frameworks, which are based on denoising diffusion probabilistic models, noise conditioned score networks, and stochastic differential equations. We further discuss the relations between diffusion models and other deep generative models, including variational auto-encoders, generative adversarial networks, energy-based models, autoregressive models and normalizing flows. Then, we introduce a multi-perspective categorization of diffusion models applied in computer vision. Finally, we illustrate the current limitations of diffusion models and envision some interesting directions for future research.
Social relations are often used to improve recommendation quality when user-item interaction data is sparse in recommender systems. Most existing social recommendation models exploit pairwise relations to mine potential user preferences. However, real-life interactions among users are very complicated and user relations can be high-order. Hypergraph provides a natural way to model complex high-order relations, while its potentials for improving social recommendation are under-explored. In this paper, we fill this gap and propose a multi-channel hypergraph convolutional network to enhance social recommendation by leveraging high-order user relations. Technically, each channel in the network encodes a hypergraph that depicts a common high-order user relation pattern via hypergraph convolution. By aggregating the embeddings learned through multiple channels, we obtain comprehensive user representations to generate recommendation results. However, the aggregation operation might also obscure the inherent characteristics of different types of high-order connectivity information. To compensate for the aggregating loss, we innovatively integrate self-supervised learning into the training of the hypergraph convolutional network to regain the connectivity information with hierarchical mutual information maximization. The experimental results on multiple real-world datasets show that the proposed model outperforms the SOTA methods, and the ablation study verifies the effectiveness of the multi-channel setting and the self-supervised task. The implementation of our model is available via //github.com/Coder-Yu/RecQ.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191