When applied to processing long text, Large Language Models (LLMs) are limited by their context window. Existing efforts to address this limitation involve training specialized architectures, and cannot be easily applied to off-the-shelf LLMs. We present Parallel Context Windows (PCW), a method that alleviates the context window restriction for any off-the-shelf LLM without further training. The key to the approach is to carve a long context into chunks (``windows''), restrict the attention mechanism to apply only within each window, and re-use the positional embeddings across the windows. Our main results test the PCW approach on in-context learning with models that range in size between 750 million and 178 billion parameters, and show substantial improvements for tasks with diverse input and output spaces. We show additional benefits in other settings where long context windows may be beneficial: multi-hop questions and retrieval-augmented question answering with multiple retrieved documents. Our results highlight Parallel Context Windows as a promising method for applying off-the-shelf LLMs in a range of settings that require long text sequences. We make our code publicly available at //github.com/ai21labs/parallel-context-windows.
相關內容
Graph Neural Networks (GNNs) have achieved tremendous success in a variety of real-world applications by relying on the fixed graph data as input. However, the initial input graph might not be optimal in terms of specific downstream tasks, because of information scarcity, noise, adversarial attacks, or discrepancies between the distribution in graph topology, features, and groundtruth labels. In this paper, we propose a bi-level optimization approach for learning the optimal graph structure via directly learning the Personalized PageRank propagation matrix as well as the downstream semi-supervised node classification simultaneously. We also explore a low-rank approximation model for further reducing the time complexity. Empirical evaluations show the superior efficacy and robustness of the proposed model over all baseline methods.
While large language models (LLMs) have demonstrated impressive performance in question-answering tasks, their performance is limited when the questions require knowledge that is not included in the model's training data and can only be acquired through direct observation or interaction with the real world. Existing methods decompose reasoning tasks through the use of modules invoked sequentially, limiting their ability to answer deep reasoning tasks. We introduce a method, Recursion based extensible LLM (REBEL), which handles open-world, deep reasoning tasks by employing automated reasoning techniques like dynamic planning and forward-chaining strategies. REBEL allows LLMs to reason via recursive problem decomposition and utilization of external tools. The tools that REBEL uses are specified only by natural language description. We further demonstrate REBEL capabilities on a set of problems that require a deeply nested use of external tools in a compositional and conversational setting.
Vector graphics are an industry-standard way to represent and share visual designs. Designers frequently source and incorporate styles from existing designs into their own work. Unfortunately, popular design tools aren't well suited for this task. We present VST, Vector Style Transfer, a novel design tool for flexibly transferring visual styles between vector graphics. The core of VST lies in leveraging automation while respecting designers' tastes and the subjectivity inherent to style transfer. In VST, designers tune a cross-design element correspondence and customize which style attributes to change. We report results from a user study in which designers used VST to control style transfer between several designs, including designs participants created with external tools beforehand. VST shows that enabling design correspondence tuning and customization is one way to support interactive, flexible style transfer. We also find that someone using VST can significantly reduce the time and work for style transfer compared to experienced designers using industry-standard tools.
Hardware-aware Neural Architecture Search (HW-NAS) is increasingly being used to design efficient deep learning architectures. An efficient and flexible search space is crucial to the success of HW-NAS. Current approaches focus on designing a macro-architecture and searching for the architecture's hyperparameters based on a set of possible values. This approach is biased by the expertise of deep learning (DL) engineers and standard modeling approaches. In this paper, we present a Grassroots Operator Search (GOS) methodology. Our HW-NAS adapts a given model for edge devices by searching for efficient operator replacement. We express each operator as a set of mathematical instructions that capture its behavior. The mathematical instructions are then used as the basis for searching and selecting efficient replacement operators that maintain the accuracy of the original model while reducing computational complexity. Our approach is grassroots since it relies on the mathematical foundations to construct new and efficient operators for DL architectures. We demonstrate on various DL models, that our method consistently outperforms the original models on two edge devices, namely Redmi Note 7S and Raspberry Pi3, with a minimum of 2.2x speedup while maintaining high accuracy. Additionally, we showcase a use case of our GOS approach in pulse rate estimation on wristband devices, where we achieve state-of-the-art performance, while maintaining reduced computational complexity, demonstrating the effectiveness of our approach in practical applications.
We propose a new score-based model with one-step sampling. Previously, score-based models were burdened with heavy computations due to iterative sampling. For substituting the iterative process, we train a standalone generator to compress all the time steps with the gradient backpropagated from the score network. In order to produce meaningful gradients for the generator, the score network is trained to simultaneously match the real data distribution and mismatch the fake data distribution. This model has the following advantages: 1) For sampling, it generates a fake image with only one step forward. 2) For training, it only needs 10 diffusion steps.3) Compared with consistency model, it is free of the ill-posed problem caused by consistency loss. On the popular CIFAR-10 dataset, our model outperforms Consistency Model and Denoising Score Matching, which demonstrates the potential of the framework. We further provide more examples on the MINIST and LSUN datasets. The code is available on GitHub.
Recent advances in Neural Radiance Fields (NeRFs) have made it possible to reconstruct and reanimate dynamic portrait scenes with control over head-pose, facial expressions and viewing direction. However, training such models assumes photometric consistency over the deformed region e.g. the face must be evenly lit as it deforms with changing head-pose and facial expression. Such photometric consistency across frames of a video is hard to maintain, even in studio environments, thus making the created reanimatable neural portraits prone to artifacts during reanimation. In this work, we propose CoDyNeRF, a system that enables the creation of fully controllable 3D portraits in real-world capture conditions. CoDyNeRF learns to approximate illumination dependent effects via a dynamic appearance model in the canonical space that is conditioned on predicted surface normals and the facial expressions and head-pose deformations. The surface normals prediction is guided using 3DMM normals that act as a coarse prior for the normals of the human head, where direct prediction of normals is hard due to rigid and non-rigid deformations induced by head-pose and facial expression changes. Using only a smartphone-captured short video of a subject for training, we demonstrate the effectiveness of our method on free view synthesis of a portrait scene with explicit head pose and expression controls, and realistic lighting effects. The project page can be found here: //shahrukhathar.github.io/2023/08/22/CoDyNeRF.html
Diagrammatic Teaching is a paradigm for robots to acquire novel skills, whereby the user provides 2D sketches over images of the scene to shape the robot's motion. In this work, we tackle the problem of teaching a robot to approach a surface and then follow cyclic motion on it, where the cycle of the motion can be arbitrarily specified by a single user-provided sketch over an image from the robot's camera. Accordingly, we introduce the \emph{Stable Diffeomorphic Diagrammatic Teaching} (SDDT) framework. SDDT models the robot's motion as an \emph{Orbitally Asymptotically Stable} (O.A.S.) dynamical system that learns to follow the user-specified sketch. This is achieved by applying a \emph{diffeomorphism}, i.e. a differentiable and invertible function, to morph a known O.A.S. system. The parameterised diffeomorphism is then optimised with respect to the Hausdorff distance between the limit cycle of our modelled system and the sketch, to produce the desired robot motion. We provide theoretical insight into the behaviour of the optimised system and also empirically evaluate SDDT, both in simulation and on a quadruped with a mounted 6-DOF manipulator. Results show that we can diagrammatically teach complex cyclic motion patterns with a high degree of accuracy.
Graph Neural Networks (GNNs) have proven to be useful for many different practical applications. However, many existing GNN models have implicitly assumed homophily among the nodes connected in the graph, and therefore have largely overlooked the important setting of heterophily, where most connected nodes are from different classes. In this work, we propose a novel framework called CPGNN that generalizes GNNs for graphs with either homophily or heterophily. The proposed framework incorporates an interpretable compatibility matrix for modeling the heterophily or homophily level in the graph, which can be learned in an end-to-end fashion, enabling it to go beyond the assumption of strong homophily. Theoretically, we show that replacing the compatibility matrix in our framework with the identity (which represents pure homophily) reduces to GCN. Our extensive experiments demonstrate the effectiveness of our approach in more realistic and challenging experimental settings with significantly less training data compared to previous works: CPGNN variants achieve state-of-the-art results in heterophily settings with or without contextual node features, while maintaining comparable performance in homophily settings.
Graph Neural Networks (GNNs) have been shown to be effective models for different predictive tasks on graph-structured data. Recent work on their expressive power has focused on isomorphism tasks and countable feature spaces. We extend this theoretical framework to include continuous features - which occur regularly in real-world input domains and within the hidden layers of GNNs - and we demonstrate the requirement for multiple aggregation functions in this context. Accordingly, we propose Principal Neighbourhood Aggregation (PNA), a novel architecture combining multiple aggregators with degree-scalers (which generalize the sum aggregator). Finally, we compare the capacity of different models to capture and exploit the graph structure via a novel benchmark containing multiple tasks taken from classical graph theory, alongside existing benchmarks from real-world domains, all of which demonstrate the strength of our model. With this work, we hope to steer some of the GNN research towards new aggregation methods which we believe are essential in the search for powerful and robust models.
Graph Neural Networks (GNNs) have recently been used for node and graph classification tasks with great success, but GNNs model dependencies among the attributes of nearby neighboring nodes rather than dependencies among observed node labels. In this work, we consider the task of inductive node classification using GNNs in supervised and semi-supervised settings, with the goal of incorporating label dependencies. Because current GNNs are not universal (i.e., most-expressive) graph representations, we propose a general collective learning approach to increase the representation power of any existing GNN. Our framework combines ideas from collective classification with self-supervised learning, and uses a Monte Carlo approach to sampling embeddings for inductive learning across graphs. We evaluate performance on five real-world network datasets and demonstrate consistent, significant improvement in node classification accuracy, for a variety of state-of-the-art GNNs.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191