Context: Roblox Studio lets millions of creators build interactive experiences by programming in a variant of Lua called Luau. The creators form a broad group, ranging from novices writing their first script to professional developers; thus, Luau must support a wide audience. As part of its efforts to support all kinds of programmers, Luau includes an optional, gradual type system and goes to great lengths to minimize false positive errors. Inquiry: Since Luau is currently being used by many creators, we want to collect data to improve the language and, in particular, the type system. The standard way to collect data is to deploy client-side telemetry; however, we cannot scrape personal data or proprietary information, which means we cannot collect source code fragments, error messages, or even filepaths. The research questions are thus about how to conduct telemetry that is not invasive and obtain insights from it about type errors. Approach: We designed and implemented a pseudonymized, randomly-sampling telemetry system for Luau. Telemetry records include a timestamp, a session id, a reason for sending, and a numeric summary of the most recent type analyses. This information lets us study type errors over time without revealing private data. We deployed the system in Roblox Studio during Spring 2023 and collected over 1.5 million telemetry records from over 340,000 sessions. Knowledge: We present several findings about Luau, all of which suggest that telemetry is an effective way to study type error pragmatics. One of the less-surprising findings is that opt-in gradual types are unpopular: there is an 100x gap between the number of untyped Luau sessions and the number of typed ones. One surprise is that the strict mode for type analysis is overly conservative about interactions with data assets. A reassuring finding is that type analysis rarely hits its internal limits on problem size. Grounding: Our findings are supported by a dataset of over 1.5 million telemetry records. The data and scripts for analyzing it are available in an artifact. Importance: Beyond the immediate benefits to Luau, our findings about types and type errors have implications for adoption and ergonomics in other gradual languages such as TypeScript, Elixir, and Typed Racket. Our telemetry design is of broad interest, as it reports on type errors without revealing sensitive information.
相關內容
With the development of large language models, multiple AIs are now made available for code generation (such as ChatGPT and StarCoder) and are adopted widely. It is often desirable to know whether a piece of code is generated by AI, and furthermore, which AI is the author. For instance, if a certain version of AI is known to generate vulnerable codes, it is particularly important to know the creator. Existing approaches are not satisfactory as watermarking codes are more challenging compared to watermarking text data, as codes can be altered with relative ease via widely-used code refactoring methods. In this work, we propose ACW (AI Code Watermarking), a novel method for watermarking AI-generated codes. The key idea of ACW is to selectively apply a set of carefully-designed semantic-preserving, idempotent code transformations, whose presence (or absence) allows us to determine the existence of the watermark. It is efficient as it requires no training or fine-tuning and works in a black-box manner. It is resilient as the watermark cannot be easily removed or tampered through common code refactoring methods. Our experimental results show that ACW is effective (i.e., achieving high accuracy, true positive rates and false positive rates) and resilient, significantly outperforming existing approaches.
Learning code representations has been the core prerequisite of many software engineering tasks such as code clone detection and code generation. State-of-the-art program representation techniques mainly utilize pre-trained language models (PLMs) such as CodeBERT. A Transformer encoder is firstly pre-trained on a large-scale code corpus to acquire general knowledge about source code. The pre-trained model is then fine-tuned on specific tasks using an amount of labeled data. However, gathering training samples for the downstream tasks can be prohibitively expensive and impractical for domain-specific languages or project-specific tasks. Besides, pre-training and downstream tasks are usually heterogeneous, which makes it difficult to fully explore the knowledge learned during pre-training. In this paper, we propose Zecoler, a zero-shot approach for learning code representations. Zecoler is built upon a pre-trained programming language model. In order to elicit knowledge from the PLMs efficiently, Zecoler casts the downstream tasks to the same form of pre-training objectives by inserting train-able prompts into the original input. These prompts can guide PLMs on how to generate better results. Subsequently, we employ the prompt tuning technique to search for the optimal prompts for PLMs automatically. This enables the representation model to efficiently fit the downstream tasks through fine-tuning on the dataset in source language domain and then reuse the pre-trained knowledge for the target domain in a zero-shot style. We evaluate Zecoler in five code intelligence tasks including code clone detection, code search, method name prediction, code summarization, and code generation. The results show that our approach significantly outperforms baseline models under the zero-shot setting.
Understanding how individuals focus and perform visual searches during collaborative tasks can help improve user engagement. Eye tracking measures provide informative cues for such understanding. This article presents A-DisETrac, an advanced analytic dashboard for distributed eye tracking. It uses off-the-shelf eye trackers to monitor multiple users in parallel, compute both traditional and advanced gaze measures in real-time, and display them on an interactive dashboard. Using two pilot studies, the system was evaluated in terms of user experience and utility, and compared with existing work. Moreover, the system was used to study how advanced gaze measures such as ambient-focal coefficient K and real-time index of pupillary activity relate to collaborative behavior. It was observed that the time a group takes to complete a puzzle is related to the ambient visual scanning behavior quantified and groups that spent more time had more scanning behavior. User experience questionnaire results suggest that their dashboard provides a comparatively good user experience.
This work develops a novel trajectory planner for human-robot handovers. The handover requirements can naturally be handled by a path-following-based model predictive controller, where the path progress serves as a progress measure of the handover. Moreover, the deviations from the path are used to follow human motion by adapting the path deviation bounds with a handover location prediction. A Gaussian process regression model, which is trained on known handover trajectories, is employed for this prediction. Experiments with a collaborative 7-DoF robotic manipulator show the effectiveness and versatility of the proposed approach.
Language models trained on internet-scale data sets have shown an impressive ability to solve problems in Natural Language Processing and Computer Vision. However, experience is showing that these models are frequently brittle in unexpected ways, and require significant scaffolding to ensure that they operate correctly in the larger systems that comprise "language-model agents." In this paper, we argue that behavior trees provide a unifying framework for combining language models with classical AI and traditional programming. We introduce Dendron, a Python library for programming language model agents using behavior trees. We demonstrate the approach embodied by Dendron in three case studies: building a chat agent, a camera-based infrastructure inspection agent for use on a mobile robot or vehicle, and an agent that has been built to satisfy safety constraints that it did not receive through instruction tuning or RLHF.
We design, to the best of our knowledge, the first differentially private (DP) stream aggregation processing system at scale. Our system -- Differential Privacy SQL Pipelines (DP-SQLP) -- is built using a streaming framework similar to Spark streaming, and is built on top of the Spanner database and the F1 query engine from Google. Towards designing DP-SQLP we make both algorithmic and systemic advances, namely, we (i) design a novel (user-level) DP key selection algorithm that can operate on an unbounded set of possible keys, and can scale to one billion keys that users have contributed, (ii) design a preemptive execution scheme for DP key selection that avoids enumerating all the keys at each triggering time, and (iii) use algorithmic techniques from DP continual observation to release a continual DP histogram of user contributions to different keys over the stream length. We empirically demonstrate the efficacy by obtaining at least $16\times$ reduction in error over meaningful baselines we consider. We implemented a streaming differentially private user impressions for Google Shopping with DP-SQLP. The streaming DP algorithms are further applied to Google Trends.
Existing prompt learning methods have shown certain capabilities in Out-of-Distribution (OOD) detection, but the lack of OOD images in the target dataset in their training can lead to mismatches between OOD images and In-Distribution (ID) categories, resulting in a high false positive rate. To address this issue, we introduce a novel OOD detection method, named 'NegPrompt', to learn a set of negative prompts, each representing a negative connotation of a given class label, for delineating the boundaries between ID and OOD images. It learns such negative prompts with ID data only, without any reliance on external outlier data. Further, current methods assume the availability of samples of all ID classes, rendering them ineffective in open-vocabulary learning scenarios where the inference stage can contain novel ID classes not present during training. In contrast, our learned negative prompts are transferable to novel class labels. Experiments on various ImageNet benchmarks show that NegPrompt surpasses state-of-the-art prompt-learning-based OOD detection methods and maintains a consistent lead in hard OOD detection in closed- and open-vocabulary classification scenarios. Code is available at //github.com/mala-lab/negprompt.
Model-free reinforcement learning (RL) can be used to learn effective policies for complex tasks, such as Atari games, even from image observations. However, this typically requires very large amounts of interaction -- substantially more, in fact, than a human would need to learn the same games. How can people learn so quickly? Part of the answer may be that people can learn how the game works and predict which actions will lead to desirable outcomes. In this paper, we explore how video prediction models can similarly enable agents to solve Atari games with fewer interactions than model-free methods. We describe Simulated Policy Learning (SimPLe), a complete model-based deep RL algorithm based on video prediction models and present a comparison of several model architectures, including a novel architecture that yields the best results in our setting. Our experiments evaluate SimPLe on a range of Atari games in low data regime of 100k interactions between the agent and the environment, which corresponds to two hours of real-time play. In most games SimPLe outperforms state-of-the-art model-free algorithms, in some games by over an order of magnitude.
Federated Learning (FL) is a decentralized machine-learning paradigm, in which a global server iteratively averages the model parameters of local users without accessing their data. User heterogeneity has imposed significant challenges to FL, which can incur drifted global models that are slow to converge. Knowledge Distillation has recently emerged to tackle this issue, by refining the server model using aggregated knowledge from heterogeneous users, other than directly averaging their model parameters. This approach, however, depends on a proxy dataset, making it impractical unless such a prerequisite is satisfied. Moreover, the ensemble knowledge is not fully utilized to guide local model learning, which may in turn affect the quality of the aggregated model. Inspired by the prior art, we propose a data-free knowledge distillation} approach to address heterogeneous FL, where the server learns a lightweight generator to ensemble user information in a data-free manner, which is then broadcasted to users, regulating local training using the learned knowledge as an inductive bias. Empirical studies powered by theoretical implications show that, our approach facilitates FL with better generalization performance using fewer communication rounds, compared with the state-of-the-art.
Knowledge graphs (KGs) serve as useful resources for various natural language processing applications. Previous KG completion approaches require a large number of training instances (i.e., head-tail entity pairs) for every relation. The real case is that for most of the relations, very few entity pairs are available. Existing work of one-shot learning limits method generalizability for few-shot scenarios and does not fully use the supervisory information; however, few-shot KG completion has not been well studied yet. In this work, we propose a novel few-shot relation learning model (FSRL) that aims at discovering facts of new relations with few-shot references. FSRL can effectively capture knowledge from heterogeneous graph structure, aggregate representations of few-shot references, and match similar entity pairs of reference set for every relation. Extensive experiments on two public datasets demonstrate that FSRL outperforms the state-of-the-art.
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