arXiv subjects
Publications and source records attributed to Yu Sun.
The Varentropy is a measure of the variability of the information content of random vector and it is invariant under affine transformations. We introduce the statistical estimate of varentropy of random vector based on the nearest neighbor graphs (distances). The asymptotic unbiasedness and L2-consistency of the estimates are established.
Reward modeling (a.k.a., preference modeling) is instrumental for aligning large language models with human preferences, particularly within the context of reinforcement learning from human feedback (RLHF). While conventional reward models (RMs) have exhibited remarkable scalability, they oft struggle with fundamental functionality such as arithmetic computation, code execution, and factual lookup. In this paper, we propose a tool-augmented preference modeling approach, named Themis, to address these limitations by empowering RMs with access to external environments, including calculators and search engines. This approach not only fosters synergy between tool utilization and reward grading but also enhances interpretive capacity and scoring reliability. Our study delves into the integration of external tools into RMs, enabling them to interact with diverse external sources and construct task-specific tool engagement and reasoning traces in an autoregressive manner. We validate our approach across a wide range of domains, incorporating seven distinct external tools. Our experimental results demonstrate a noteworthy overall improvement of 17.7% across eight tasks in preference ranking. Furthermore, our approach outperforms Gopher 280B by 7.3% on TruthfulQA task in zero-shot evaluation. In human evaluations, RLHF trained with Themis attains an average win rate of 32% when compared to baselines across four distinct tasks. Additionally, we provide a comprehensive collection of tool-related RM datasets, incorporating data from seven distinct tool APIs, totaling 15,000 instances. We have made the code, data, and model checkpoints publicly available to facilitate and inspire further research advancements\footnote{\url{https://github.com/ernie-research/Tool-Augmented-Reward-Model}}.
Many recent efforts augment language models with retrieval, by adding retrieved data to the input context. For this approach to succeed, the retrieved data must be added at both training and test time. Moreover, as input length grows linearly with the size of retrieved data, cost in computation and memory grows quadratically for modern Transformers. To avoid these complications, we simply fine-tune the model on retrieved data at test time, using its standard training setup. We build a large-scale distributed index based on text embeddings of the Pile dataset. For each test input, our system retrieves its neighbors and fine-tunes the model on their text. Surprisingly, retrieving and training on as few as 20 neighbors, each for only one gradient iteration, drastically improves performance across more than 20 language modeling tasks in the Pile. For example, test-time training with nearest neighbors significantly narrows the performance gap between a small GPT-2 and a GPT-Neo model more than 10 times larger. Sufficient index quality and size, however, are necessary. Our work establishes a first baseline of test-time training for language modeling.
The inherent probabilistic nature of Large Language Models (LLMs) introduces an element of unpredictability, raising concerns about potential discrepancies in their output. This paper introduces an innovative approach aims to generate correct and optimal robotic task plans for diverse real-world demands and scenarios. LLMs have been used to generate task plans, but they are unreliable and may contain wrong, questionable, or high-cost steps. The proposed approach uses LLM to generate a number of task plans as trees and amalgamates them into a graph by removing questionable paths. Then an optimal task tree can be retrieved to circumvent questionable and high-cost nodes, thereby improving planning accuracy and execution efficiency. The approach is further improved by incorporating a large knowledge network. Leveraging GPT-4 further, the high-level task plan is converted into a low-level Planning Domain Definition Language (PDDL) plan executable by a robot. Evaluation results highlight the superior accuracy and efficiency of our approach compared to previous methodologies in the field of task planning.
We reformulate the problem of supervised learning as learning to learn with two nested loops (i.e. learning problems). The inner loop learns on each individual instance with self-supervision before final prediction. The outer loop learns the self-supervised task used by the inner loop, such that its final prediction improves. Our inner loop turns out to be equivalent to linear attention when the inner-loop learner is only a linear model, and to self-attention when it is a kernel estimator. For practical comparison with linear or self-attention layers, we replace each of them in a transformer with an inner loop, so our outer loop is equivalent to training the architecture. When each inner-loop learner is a neural network, our approach vastly outperforms transformers with linear attention on ImageNet from 224 x 224 raw pixels in both accuracy and FLOPs, while (regular) transformers cannot run.
Earth system tipping elements have been predominantly investigated for their potential to trigger \textit{negative} ecological, climatic, and societal shifts. Yet, an overlooked but seminal avenue exists in the form of \textit{positive} tipping phenomena, whose underlying mechanisms and benefits remain largely underexplored. To bridge this gap, our research introduces a fundamental percolation-based framework to assess the criticality and resilience of planetary terrestrial vegetation systems. Leveraging high-resolution satellite data, we focus on greening-induced positive tipping dynamics driven by global warming. We feature the Qinghai-Tibetan Plateau (QTP) and the Sahel region as contrasting yet analogous case studies. Our analysis uncovers an intriguing phenomenon where vegetation fragmentation aligns with a percolation threshold, exhibiting a scale-invariant pattern characterized by nearly perfect power laws with three critical exponents. Remarkably, contrary to conventional destructive tipping elements, these regions act as favorable tipping elements, transitioning from fragmented to cohesive vegetation patterns due to anthropogenic climate change and afforestation efforts. Furthermore, we propose an \textit{optimal resilience enhancement model} to reinforce vegetation robustness while minimizing socio-economic costs. This study provides valuable insights into the favorable aspects of tipping elements under climate change and offers effective strategies for enhancing ecological resilience against environmental threats.
Large language models (LLMs) are increasingly pivotal in a wide range of natural language processing tasks. Access to pre-trained models, courtesy of the open-source community, has made it possible to adapt these models to specific applications for enhanced performance. However, the substantial resources required for training these models necessitate efficient solutions. This paper introduces CoLLiE, an efficient library that facilitates collaborative training of large language models using 3D parallelism, parameter-efficient fine-tuning (PEFT) methods, and optimizers such as Lion, Adan, Sophia, LOMO and AdaLomo. With its modular design and comprehensive functionality, CoLLiE offers a balanced blend of efficiency, ease of use, and customization. CoLLiE has proven superior training efficiency in comparison with prevalent solutions in pre-training and fine-tuning scenarios. Furthermore, we provide an empirical evaluation of the correlation between model size and GPU memory consumption under different optimization methods, as well as an analysis of the throughput. Lastly, we carry out a comprehensive comparison of various optimizers and PEFT methods within the instruction-tuning context. CoLLiE is available at https://github.com/OpenLMLab/collie.
Significant wave height is one of the most important parameters characterizing ocean waves, and accurate numerical ocean wave forecasting is crucial for coastal protection and shipping. However, due to the randomness and nonlinearity of the wind fields that generate ocean waves and the complex interaction between wave and wind fields, current forecasts of numerical ocean waves have biases. In this study, a spatiotemporal deep-learning method was employed to correct gridded SWH forecasts from the ECMWF-IFS. This method was built on the trajectory gated recurrent unit deep neural network,and it conducts real-time rolling correction for the 0-240h SWH forecasts from ECMWF-IFS. The correction model is co-driven by wave and wind fields, providing better results than those based on wave fields alone. A novel pixel-switch loss function was developed. The pixel-switch loss function can dynamically fine-tune the pre-trained correction model, focusing on pixels with large biases in SWH forecasts. According to the seasonal characteristics of SWH, four correction models were constructed separately, for spring, summer, autumn, and winter. The experimental results show that, compared with the original ECMWF SWH predictions, the correction was most effective in spring, when the mean absolute error decreased by 12.972~46.237%. Although winter had the worst performance, the mean absolute error decreased by 13.794~38.953%. The corrected results improved the original ECMWF SWH forecasts under both normal and extreme weather conditions, indicating that our SWH correction model is robust and generalizable.
Although the estimation of 3D human pose and shape (HPS) is rapidly progressing, current methods still cannot reliably estimate moving humans in global coordinates, which is critical for many applications. This is particularly challenging when the camera is also moving, entangling human and camera motion. To address these issues, we adopt a novel 5D representation (space, time, and identity) that enables end-to-end reasoning about people in scenes. Our method, called TRACE, introduces several novel architectural components. Most importantly, it uses two new "maps" to reason about the 3D trajectory of people over time in camera, and world, coordinates. An additional memory unit enables persistent tracking of people even during long occlusions. TRACE is the first one-stage method to jointly recover and track 3D humans in global coordinates from dynamic cameras. By training it end-to-end, and using full image information, TRACE achieves state-of-the-art performance on tracking and HPS benchmarks. The code and dataset are released for research purposes.
This paper investigates a generalized hyperbolic circle packing (including circles, horocycles or hypercycles) with respect to the total geodesic curvatures on the surface with boundary. We mainly focus on the existence and rigidity of circle packing whose contact graph is the $1$-skeleton of a finite polygonal cellular decomposition, which is analogous to the construction of Bobenko and Springborn [4]. Motivated by Colin de Verdière's method [6], we introduce the variational principle for generalized hyperbolic circle packings on polygons. By analyzing limit behaviours of generalized circle packings on polygons, we give an existence and rigidity for the generalized hyperbolic circle packing with conical singularities regarding the total geodesic curvature on each vertex of the contact graph. As a consequence, we introduce the combinatoral Ricci flow to find a desired circle packing with a prescribed total geodesic curvature on each vertex of the contact graph.
This study aims to analyze the service and return landing areas in badminton men's double, based on data extracted from 20 badminton matches. We find that most services land near the center-line, while returns tend to land in the crossing areas of the serving team's court. Using generalized logit models, we are able to predict the return landing area based on features of the service and return round. We find that the direction of the service and the footwork and grip of the receiver could indicate his intended return landing area. Additionally, we discover that servers tend to intercept in specific areas based on their serving position. Our results offer valuable insights into the strategic decisions made by players in the service and return of a badminton rally.
The device fingerprinting technique extracts fingerprints based on the hardware characteristics of the device to identify the device. The primary goal of device fingerprinting is to accurately and uniquely identify a device, which requires the generated device fingerprints to have good stability to achieve long-term tracking of the target device. However, the fingerprints generated by some existing fingerprinting technologies are not stable enough or change frequently, making it impossible to track the target device for a long time. In this paper, we present FPHammer, a novel DRAM-based fingerprinting technique. The device fingerprint generated by our technique has high stability and can be used to track the device for a long time. We leverage the Rowhammer technique to repeatedly and quickly access a row in DRAM to get bit flips in its adjacent row. We then construct a physical fingerprint of the device based on the locations of the collected bit flips. The evaluation results of the uniqueness and reliability of the physical fingerprint show that it can be used to distinguish devices with the same hardware and software configuration. The experimental results on device identification demonstrate that the physical fingerprints engendered by our innovative technique are inherently linked to the entirety of the device rather than just the DRAM module. Even if the device modifies software-level parameters such as MAC address and IP address or even reinstalls the operating system, we can accurately identify the target device. This demonstrates that FPHammer can generate stable fingerprints that are not affected by software layer parameters.
In recent years, the burgeoning interest in diffusion models has led to significant advances in image and speech generation. Nevertheless, the direct synthesis of music waveforms from unrestricted textual prompts remains a relatively underexplored domain. In response to this lacuna, this paper introduces a pioneering contribution in the form of a text-to-waveform music generation model, underpinned by the utilization of diffusion models. Our methodology hinges on the innovative incorporation of free-form textual prompts as conditional factors to guide the waveform generation process within the diffusion model framework. Addressing the challenge of limited text-music parallel data, we undertake the creation of a dataset by harnessing web resources, a task facilitated by weak supervision techniques. Furthermore, a rigorous empirical inquiry is undertaken to contrast the efficacy of two distinct prompt formats for text conditioning, namely, music tags and unconstrained textual descriptions. The outcomes of this comparative analysis affirm the superior performance of our proposed model in terms of enhancing text-music relevance. Finally, our work culminates in a demonstrative exhibition of the excellent capabilities of our model in text-to-music generation. We further demonstrate that our generated music in the waveform domain outperforms previous works by a large margin in terms of diversity, quality, and text-music relevance.
Task planning for robotic cooking involves generating a sequence of actions for a robot to prepare a meal successfully. This paper introduces a novel task tree generation pipeline producing correct planning and efficient execution for cooking tasks. Our method first uses a large language model (LLM) to retrieve recipe instructions and then utilizes a fine-tuned GPT-3 to convert them into a task tree, capturing sequential and parallel dependencies among subtasks. The pipeline then mitigates the uncertainty and unreliable features of LLM outputs using task tree retrieval. We combine multiple LLM task tree outputs into a graph and perform a task tree retrieval to avoid questionable nodes and high-cost nodes to improve planning correctness and improve execution efficiency. Our evaluation results show its superior performance compared to previous works in task planning accuracy and efficiency.
Recent 2D-to-3D human pose estimation (HPE) utilizes temporal consistency across sequences to alleviate the depth ambiguity problem but ignore the action related prior knowledge hidden in the pose sequence. In this paper, we propose a plug-and-play module named Action Prompt Module (APM) that effectively mines different kinds of action clues for 3D HPE. The highlight is that, the mining scheme of APM can be widely adapted to different frameworks and bring consistent benefits. Specifically, we first present a novel Action-related Text Prompt module (ATP) that directly embeds action labels and transfers the rich language information in the label to the pose sequence. Besides, we further introduce Action-specific Pose Prompt module (APP) to mine the position-aware pose pattern of each action, and exploit the correlation between the mined patterns and input pose sequence for further pose refinement. Experiments show that APM can improve the performance of most video-based 2D-to-3D HPE frameworks by a large margin.
Picking up multiple objects at once is a grasping skill that makes a human worker efficient in many domains. This paper presents a system to pick a requested number of objects by only picking once (OPO). The proposed Only-Pick-Once System (OPOS) contains several graph-based algorithms that convert the layout of objects into a graph, cluster nodes in the graph, rank and select candidate clusters based on their topology. OPOS also has a multi-object picking predictor based on a convolutional neural network for estimating how many objects would be picked up with a given gripper location and orientation. This paper presents four evaluation metrics and three protocols to evaluate the proposed OPOS. The results show OPOS has very high success rates for two and three objects when only picking once. Using OPOS can significantly outperform two to three times single object picking in terms of efficiency. The results also show OPOS can generalize to unseen size and shape objects.
Within the realm of service robotics, researchers have placed a great amount of effort into learning, understanding, and representing motions as manipulations for task execution by robots. The task of robot learning and problem-solving is very broad, as it integrates a variety of tasks such as object detection, activity recognition, task/motion planning, localization, knowledge representation and retrieval, and the intertwining of perception/vision and machine learning techniques. In this paper, we solely focus on knowledge representations and notably how knowledge is typically gathered, represented, and reproduced to solve problems as done by researchers in the past decades. In accordance with the definition of knowledge representations, we discuss the key distinction between such representations and useful learning models that have extensively been introduced and studied in recent years, such as machine learning, deep learning, probabilistic modelling, and semantic graphical structures. Along with an overview of such tools, we discuss the problems which have existed in robot learning and how they have been built and used as solutions, technologies or developments (if any) which have contributed to solving them. Finally, we discuss key principles that should be considered when designing an effective knowledge representation.
Software engineers working with the same programming language (PL) may speak different natural languages (NLs) and vice versa, erecting huge barriers to communication and working efficiency. Recent studies have demonstrated the effectiveness of generative pre-training in computer programs, yet they are always English-centric. In this work, we step towards bridging the gap between multilingual NLs and multilingual PLs for large language models (LLMs). We release ERNIE-Code, a unified pre-trained language model for 116 NLs and 6 PLs. We employ two methods for universal cross-lingual pre-training: span-corruption language modeling that learns patterns from monolingual NL or PL; and pivot-based translation language modeling that relies on parallel data of many NLs and PLs. Extensive results show that ERNIE-Code outperforms previous multilingual LLMs for PL or NL across a wide range of end tasks of code intelligence, including multilingual code-to-text, text-to-code, code-to-code, and text-to-text generation. We further show its advantage of zero-shot prompting on multilingual code summarization and text-to-text translation. We release our code and pre-trained checkpoints.