The HuggingFace Datasets Hub hosts thousands of datasets, offering exciting opportunities for language model training and evaluation. However, datasets for a specific task type often have different schemas, making harmonization challenging. Multi-task training or evaluation necessitates manual work to fit data into task templates. Several initiatives independently tackle this issue by releasing harmonized datasets or providing harmonization codes to preprocess datasets into a consistent format. We identify patterns across previous preprocessing efforts, such as column name mapping and extracting specific sub-fields from structured data in a column. We then propose a structured annotation framework that ensures our annotations are fully exposed and not hidden within unstructured code. We release a dataset annotation framework and dataset annotations for more than 500 English tasks\footnote{\url{https://github.com/sileod/tasksource}}. These annotations include metadata, such as the names of columns to be used as input or labels for all datasets, which can save time for future dataset preprocessing, regardless of whether our framework is utilized. We fine-tune a multi-task text encoder on all tasksource tasks, outperforming every publicly available text encoder of comparable size in an external evaluation.

It has been commonly observed that a teacher model with superior performance does not necessarily result in a stronger student, highlighting a discrepancy between current teacher training practices and effective knowledge transfer. In order to enhance the guidance of the teacher training process, we introduce the concept of distillation influence to determine the impact of distillation from each training sample on the student's generalization ability. In this paper, we propose Learning Good Teacher Matters (LGTM), an efficient training technique for incorporating distillation influence into the teacher's learning process. By prioritizing samples that are likely to enhance the student's generalization ability, our LGTM outperforms 10 common knowledge distillation baselines on 6 text classification tasks in the GLUE benchmark.

Humans use semantic concepts such as spatial relations between objects to describe scenes and communicate tasks such as "Put the tea to the right of the cup" or "Move the plate between the fork and the spoon." Just as children, assistive robots must be able to learn the sub-symbolic meaning of such concepts from human demonstrations and instructions. We address the problem of incrementally learning geometric models of spatial relations from few demonstrations collected online during interaction with a human. Such models enable a robot to manipulate objects in order to fulfill desired spatial relations specified by verbal instructions. At the start, we assume the robot has no geometric model of spatial relations. Given a task as above, the robot requests the user to demonstrate the task once in order to create a model from a single demonstration, leveraging cylindrical probability distribution as generative representation of spatial relations. We show how this model can be updated incrementally with each new demonstration without access to past examples in a sample-efficient way using incremental maximum likelihood estimation, and demonstrate the approach on a real humanoid robot.

Recent artificial intelligence (AI) systems have reached milestones in "grand challenges" ranging from Go to protein-folding. The capability to retrieve medical knowledge, reason over it, and answer medical questions comparably to physicians has long been viewed as one such grand challenge. Large language models (LLMs) have catalyzed significant progress in medical question answering; Med-PaLM was the first model to exceed a "passing" score in US Medical Licensing Examination (USMLE) style questions with a score of 67.2% on the MedQA dataset. However, this and other prior work suggested significant room for improvement, especially when models' answers were compared to clinicians' answers. Here we present Med-PaLM 2, which bridges these gaps by leveraging a combination of base LLM improvements (PaLM 2), medical domain finetuning, and prompting strategies including a novel ensemble refinement approach. Med-PaLM 2 scored up to 86.5% on the MedQA dataset, improving upon Med-PaLM by over 19% and setting a new state-of-the-art. We also observed performance approaching or exceeding state-of-the-art across MedMCQA, PubMedQA, and MMLU clinical topics datasets. We performed detailed human evaluations on long-form questions along multiple axes relevant to clinical applications. In pairwise comparative ranking of 1066 consumer medical questions, physicians preferred Med-PaLM 2 answers to those produced by physicians on eight of nine axes pertaining to clinical utility (p < 0.001). We also observed significant improvements compared to Med-PaLM on every evaluation axis (p < 0.001) on newly introduced datasets of 240 long-form "adversarial" questions to probe LLM limitations. While further studies are necessary to validate the efficacy of these models in real-world settings, these results highlight rapid progress towards physician-level performance in medical question answering.

This paper evaluates the capability of two state-of-the-art artificial intelligence (AI) models, GPT-3.5 and Bard, in generating Java code given a function description. We sourced the descriptions from CodingBat.com, a popular online platform that provides practice problems to learn programming. We compared the Java code generated by both models based on correctness, verified through the platform's own test cases. The results indicate clear differences in the capabilities of the two models. GPT-3.5 demonstrated superior performance, generating correct code for approximately 90.6% of the function descriptions, whereas Bard produced correct code for 53.1% of the functions. While both models exhibited strengths and weaknesses, these findings suggest potential avenues for the development and refinement of more advanced AI-assisted code generation tools. The study underlines the potential of AI in automating and supporting aspects of software development, although further research is required to fully realize this potential.

We propose a method that allows to develop shared understanding between two agents for the purpose of performing a task that requires cooperation. Our method focuses on efficiently establishing successful task-oriented communication in an open multi-agent system, where the agents do not know anything about each other and can only communicate via grounded interaction. The method aims to assist researchers that work on human-machine interaction or scenarios that require a human-in-the-loop, by defining interaction restrictions and efficiency metrics. To that end, we point out the challenges and limitations of such a (diverse) setup, while also restrictions and requirements which aim to ensure that high task performance truthfully reflects the extent to which the agents correctly understand each other. Furthermore, we demonstrate a use-case where our method can be applied for the task of cooperative query answering. We design the experiments by modifying an established ontology alignment benchmark. In this example, the agents want to query each other, while representing different databases, defined in their own ontologies that contain different and incomplete knowledge. Grounded interaction here has the form of examples that consists of common instances, for which the agents are expected to have similar knowledge. Our experiments demonstrate successful communication establishment under the required restrictions, and compare different agent policies that aim to solve the task in an efficient manner.

A symmetry of a state $\vert \psi \rangle$ is a unitary operator of which $\vert \psi \rangle$ is an eigenvector. When $\vert \psi \rangle$ is an unknown state supplied by a black-box oracle, the state's symmetries provide key physical insight into the quantum system; symmetries also boost many crucial quantum learning techniques. In this paper, we develop a variational hybrid quantum-classical learning scheme to systematically probe for symmetries of $\vert \psi \rangle$ with no a priori assumptions about the state. This procedure can be used to learn various symmetries at the same time. In order to avoid re-learning already known symmetries, we introduce an interactive protocol with a classical deep neural net. The classical net thereby regularizes against repetitive findings and allows our algorithm to terminate empirically with all possible symmetries found. Our scheme can be implemented efficiently on average with non-local SWAP gates; we also give a less efficient algorithm with only local operations, which may be more appropriate for current noisy quantum devices. We simulate our algorithm on representative families of states, including cluster states and ground states of Rydberg and Ising Hamiltonians. We also find that the numerical query complexity scales well with qubit size.

The issue of group fairness in machine learning models, where certain sub-populations or groups are favored over others, has been recognized for some time. While many mitigation strategies have been proposed in centralized learning, many of these methods are not directly applicable in federated learning, where data is privately stored on multiple clients. To address this, many proposals try to mitigate bias at the level of clients before aggregation, which we call locally fair training. However, the effectiveness of these approaches is not well understood. In this work, we investigate the theoretical foundation of locally fair training by studying the relationship between global model fairness and local model fairness. Additionally, we prove that for a broad class of fairness metrics, the global model's fairness can be obtained using only summary statistics from local clients. Based on that, we propose a globally fair training algorithm that directly minimizes the penalized empirical loss. Real-data experiments demonstrate the promising performance of our proposed approach for enhancing fairness while retaining high accuracy compared to locally fair training methods.

The prevalence of memes on social media has created the need to sentiment analyze their underlying meanings for censoring harmful content. Meme censoring systems by machine learning raise the need for a semi-supervised learning solution to take advantage of the large number of unlabeled memes available on the internet and make the annotation process less challenging. Moreover, the approach needs to utilize multimodal data as memes' meanings usually come from both images and texts. This research proposes a multimodal semi-supervised learning approach that outperforms other multimodal semi-supervised learning and supervised learning state-of-the-art models on two datasets, the Multimedia Automatic Misogyny Identification and Hateful Memes dataset. Building on the insights gained from Contrastive Language-Image Pre-training, which is an effective multimodal learning technique, this research introduces SemiMemes, a novel training method that combines auto-encoder and classification task to make use of the resourceful unlabeled data.

The pervasive uncertainty and dynamic nature of real-world environments present significant challenges for the widespread implementation of machine-driven Intelligent Decision-Making (IDM) systems. Consequently, IDM should possess the ability to continuously acquire new skills and effectively generalize across a broad range of applications. The advancement of Artificial General Intelligence (AGI) that transcends task and application boundaries is critical for enhancing IDM. Recent studies have extensively investigated the Transformer neural architecture as a foundational model for various tasks, including computer vision, natural language processing, and reinforcement learning. We propose that a Foundation Decision Model (FDM) can be developed by formulating diverse decision-making tasks as sequence decoding tasks using the Transformer architecture, offering a promising solution for expanding IDM applications in complex real-world situations. In this paper, we discuss the efficiency and generalization improvements offered by a foundation decision model for IDM and explore its potential applications in multi-agent game AI, production scheduling, and robotics tasks. Lastly, we present a case study demonstrating our FDM implementation, DigitalBrain (DB1) with 1.3 billion parameters, achieving human-level performance in 870 tasks, such as text generation, image captioning, video game playing, robotic control, and traveling salesman problems. As a foundation decision model, DB1 represents an initial step toward more autonomous and efficient real-world IDM applications.