The accurate predictions and principled uncertainty measures provided by GP regression incur O(n^3) cost which is prohibitive for modern-day large-scale applications. This has motivated extensive work on computationally efficient approximations. We introduce a new perspective by exploring robustness properties and limiting behaviour of GP nearest-neighbour (GPnn) prediction. We demonstrate through theory and simulation that as the data-size n increases, accuracy of estimated parameters and GP model assumptions become increasingly irrelevant to GPnn predictive accuracy. Consequently, it is sufficient to spend small amounts of work on parameter estimation in order to achieve high MSE accuracy, even in the presence of gross misspecification. In contrast, as n tends to infinity, uncertainty calibration and NLL are shown to remain sensitive to just one parameter, the additive noise-variance; but we show that this source of inaccuracy can be corrected for, thereby achieving both well-calibrated uncertainty measures and accurate predictions at remarkably low computational cost. We exhibit a very simple GPnn regression algorithm with stand-out performance compared to other state-of-the-art GP approximations as measured on large UCI datasets. It operates at a small fraction of those other methods' training costs, for example on a basic laptop taking about 30 seconds to train on a dataset of size n = 1.6 x 10^6.
相關內容
機器學習系(xi)統(tong)設計系(xi)統(tong)評估標準(zhun)
相互獨立的
·
It is a manuscript for results about entropic central limit theorem for independent sum under finite Poincar\'e constant conditions.
Motivated by the success of Transformers when applied to sequences of discrete symbols, token-based world models (TBWMs) were recently proposed as sample-efficient methods. In TBWMs, the world model consumes agent experience as a language-like sequence of tokens, where each observation constitutes a sub-sequence. However, during imagination, the sequential token-by-token generation of next observations results in a severe bottleneck, leading to long training times, poor GPU utilization, and limited representations. To resolve this bottleneck, we devise a novel Parallel Observation Prediction (POP) mechanism. POP augments a Retentive Network (RetNet) with a novel forward mode tailored to our reinforcement learning setting. We incorporate POP in a novel TBWM agent named REM (Retentive Environment Model), showcasing a 15.4x faster imagination compared to prior TBWMs. REM attains superhuman performance on 12 out of 26 games of the Atari 100K benchmark, while training in less than 12 hours. Our code is available at \url{//github.com/leor-c/REM}.
An algorithm effects a causal representation of relations between features and labels in the human's perception. Such a representation might conflict with the human's prior belief. Explanations can direct the human's attention to the conflicting feature and away from other relevant features. This leads to causal overattribution and may adversely affect the human's information processing. In a field experiment we implemented an XGBoost-trained model as a decision-making aid for counselors at a public employment service to predict candidates' risk of long-term unemployment. The treatment group of counselors was also provided with SHAP. The results show that the quality of the human's decision-making is worse when a feature on which the human holds a conflicting prior belief is displayed as part of the explanation.
We explore how much knowing a parametric restriction on propensity scores improves semiparametric efficiency bounds in the potential outcome framework. For stratified propensity scores, considered as a parametric model, we derive explicit formulas for the efficiency gain from knowing how the covariate space is split. Based on these, we find that the efficiency gain decreases as the partition of the stratification becomes finer. For general parametric models, where it is hard to obtain explicit representations of efficiency bounds, we propose a novel framework that enables us to see whether knowing a parametric model is valuable in terms of efficiency even when it is very high-dimensional. In addition to the intuitive fact that knowing the parametric model does not help much if it is sufficiently flexible, we reveal that the efficiency gain can be nearly zero even though the parametric assumption significantly restricts the space of possible propensity scores.
Quantum devices use qubits to represent information, which allows them to exploit important properties from quantum physics, specifically superposition and entanglement. As a result, quantum computers have the potential to outperform the most advanced classical computers. In recent years, quantum algorithms have shown hints of this promise, and many algorithms have been proposed for the quantum domain. There are two key hurdles to solving difficult real-world problems on quantum computers. The first is on the hardware front -- the number of qubits in the most advanced quantum systems is too small to make the solution of large problems practical. The second involves the algorithms themselves -- as quantum computers use qubits, the algorithms that work there are fundamentally different from those that work on traditional computers. As a result of these constraints, research has focused on developing approaches to solve small versions of problems as proofs of concept -- recognizing that it would be possible to scale these up once quantum devices with enough qubits become available. Our objective in this paper is along the same lines. We present a quantum approach to solve a well-studied problem in the context of data sharing. This heuristic uses the well-known Quantum Approximate Optimization Algorithm (QAOA). We present results on experiments involving small datasets to illustrate how the problem could be solved using quantum algorithms. The results show that the method has potential and provide answers close to optimal. At the same time, we realize there are opportunities for improving the method further.
Bayesian optimization has been successfully applied to optimize black-box functions where the number of evaluations is severely limited. However, in many real-world applications, it is hard or impossible to know in advance which designs are feasible due to some physical or system limitations. These issues lead to an even more challenging problem of optimizing an unknown function with unknown constraints. In this paper, we observe that in such scenarios optimal solution typically lies on the boundary between feasible and infeasible regions of the design space, making it considerably more difficult than that with interior optima. Inspired by this observation, we propose BE-CBO, a new Bayesian optimization method that efficiently explores the boundary between feasible and infeasible designs. To identify the boundary, we learn the constraints with an ensemble of neural networks that outperform the standard Gaussian Processes for capturing complex boundaries. Our method demonstrates superior performance against state-of-the-art methods through comprehensive experiments on synthetic and real-world benchmarks.
Robust fine-tuning aims to ensure performance on out-of-distribution (OOD) samples, which is sometimes compromised by pursuing adaptation on in-distribution (ID) samples. However, another criterion for reliable machine learning -- confidence calibration has been overlooked despite its increasing demand for real-world high-stakes applications, e.g., autonomous driving. We raise concerns about the calibration of fine-tuned vision-language models (VLMs) under distribution shift by showing that naive fine-tuning and even state-of-the-art robust fine-tuning hurt the calibration of pre-trained VLMs, especially on OOD datasets. We first show the OOD calibration error is bounded from above with ID calibration errors and domain discrepancy between ID and OOD. From this analysis, we propose CaRot, a calibrated robust fine-tuning method that incentivizes ID calibration and robust prediction across domains to reduce the upper bound of OOD calibration error. Extensive experiments on three types of distribution shifts (natural, synthetic, and adversarial) on ImageNet-1K classification demonstrate the effectiveness of CaRot across diverse environments. We justify the empirical success of CaRot through our theoretical analysis.
The total energy cost of computing activities is steadily increasing and projections indicate that it will be one of the dominant global energy consumers in the coming decades. However, perhaps due to its relative youth, the video game sector has not yet developed the same level of environmental awareness as other computing technologies despite the estimated three billion regular video game players in the world. This work evaluates the energy consumption of the most widely used industry-scale video game engines: Unity and Unreal Engine. Specifically, our work uses three scenarios representing relevant aspects of video games (Physics, Statics Meshes, and Dynamic Meshes) to compare the energy consumption of the engines. The aim is to determine the influence of using each of the two engines on energy consumption. Our research has confirmed significant differences in the energy consumption of video game engines: 351% in Physics in favor of Unity, 17% in Statics Meshes in favor of Unity, and 26% in Dynamic Meshes in favor of Unreal Engine. These results represent an opportunity for worldwide potential savings of at least 51 TWh per year, equivalent to the annual consumption of nearly 13 million European households, that might encourage a new branch of research on energy-efficient video game engines.
Probabilistic forecasting of irregularly sampled multivariate time series with missing values is an important problem in many fields, including health care, astronomy, and climate. State-of-the-art methods for the task estimate only marginal distributions of observations in single channels and at single timepoints, assuming a fixed-shape parametric distribution. In this work, we propose a novel model, ProFITi, for probabilistic forecasting of irregularly sampled time series with missing values using conditional normalizing flows. The model learns joint distributions over the future values of the time series conditioned on past observations and queried channels and times, without assuming any fixed shape of the underlying distribution. As model components, we introduce a novel invertible triangular attention layer and an invertible non-linear activation function on and onto the whole real line. We conduct extensive experiments on four datasets and demonstrate that the proposed model provides $4$ times higher likelihood over the previously best model.
Language models (LMs) have already demonstrated remarkable abilities in understanding and generating both natural and formal language. Despite these advances, their integration with real-world environments such as large-scale knowledge bases (KBs) remains an underdeveloped area, affecting applications such as semantic parsing and indulging in "hallucinated" information. This paper is an experimental investigation aimed at uncovering the robustness challenges that LMs encounter when tasked with knowledge base question answering (KBQA). The investigation covers scenarios with inconsistent data distribution between training and inference, such as generalization to unseen domains, adaptation to various language variations, and transferability across different datasets. Our comprehensive experiments reveal that even when employed with our proposed data augmentation techniques, advanced small and large language models exhibit poor performance in various dimensions. While the LM is a promising technology, the robustness of the current form in dealing with complex environments is fragile and of limited practicality because of the data distribution issue. This calls for future research on data collection and LM learning paradims.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191