Performance of machine learning models may differ between training and deployment for many reasons. For instance, model performance can change between environments due to changes in data quality, observing a different population than the one in training, or changes in the relationship between labels and features. These manifest as changes to the underlying data generating mechanisms, and thereby result in distribution shifts across environments. Attributing performance changes to specific shifts, such as covariate or concept shifts, is critical for identifying sources of model failures, and for taking mitigating actions that ensure robust models. In this work, we introduce the problem of attributing performance differences between environments to shifts in the underlying data generating mechanisms. We formulate the problem as a cooperative game and derive an importance weighting method for computing the value of a coalition (or a set) of distributions. The contribution of each distribution to the total performance change is then quantified as its Shapley value. We demonstrate the correctness and utility of our method on two synthetic datasets and two real-world case studies, showing its effectiveness in attributing performance changes to a wide range of distribution shifts.
相關內容
The increasing reliance on ML models in high-stakes tasks has raised a major concern on fairness violations. Although there has been a surge of work that improves algorithmic fairness, most of them are under the assumption of an identical training and test distribution. In many real-world applications, however, such an assumption is often violated as previously trained fair models are often deployed in a different environment, and the fairness of such models has been observed to collapse. In this paper, we study how to transfer model fairness under distribution shifts, a widespread issue in practice. We conduct a fine-grained analysis of how the fair model is affected under different types of distribution shifts and find that domain shifts are more challenging than subpopulation shifts. Inspired by the success of self-training in transferring accuracy under domain shifts, we derive a sufficient condition for transferring group fairness. Guided by it, we propose a practical algorithm with a fair consistency regularization as the key component. A synthetic dataset benchmark, which covers all types of distribution shifts, is deployed for experimental verification of the theoretical findings. Experiments on synthetic and real datasets including image and tabular data demonstrate that our approach effectively transfers fairness and accuracy under various distribution shifts.
Surrogate models have shown to be an extremely efficient aid in solving engineering problems that require repeated evaluations of an expensive computational model. They are built by sparsely evaluating the costly original model and have provided a way to solve otherwise intractable problems. A crucial aspect in surrogate modelling is the assumption of smoothness and regularity of the model to approximate. This assumption is however not always met in reality. For instance in civil or mechanical engineering, some models may present discontinuities or non-smoothness, e.g., in case of instability patterns such as buckling or snap-through. Building a single surrogate model capable of accounting for these fundamentally different behaviors or discontinuities is not an easy task. In this paper, we propose a three-stage approach for the approximation of non-smooth functions which combines clustering, classification and regression. The idea is to split the space following the localized behaviors or regimes of the system and build local surrogates that are eventually assembled. A sequence of well-known machine learning techniques are used: Dirichlet process mixtures models (DPMM), support vector machines and Gaussian process modelling. The approach is tested and validated on two analytical functions and a finite element model of a tensile membrane structure.
Supervised deep learning is most commonly applied to difficult problems defined on large and often extensively curated datasets. Here we demonstrate the ability of deep representation learning to address problems of classification and regression from small and poorly formed tabular datasets by encoding input information as abstracted sequences composed of a fixed number of characters per input field. We find that small models have sufficient capacity for approximation of various functions and achieve record classification benchmark accuracy. Such models are shown to form useful embeddings of various input features in their hidden layers, even if the learned task does not explicitly require knowledge of those features. These models are also amenable to input attribution, allowing for an estimation of the importance of each input element to the model output as well as of which inputs features are effectively embedded in the model. We present a proof-of-concept for the application of small language models to mixed tabular data without explicit feature engineering, cleaning, or preprocessing, relying on the model to perform these tasks as part of the representation learning process.
Recently, the use of synthetic training data has been on the rise as it offers correctly labelled datasets at a lower cost. The downside of this technique is that the so-called domain gap between the real target images and synthetic training data leads to a decrease in performance. In this paper, we attempt to provide a holistic overview of how to use synthetic data for object detection. We analyse aspects of generating the data as well as techniques used to train the models. We do so by devising a number of experiments, training models on the Dataset of Industrial Metal Objects (DIMO). This dataset contains both real and synthetic images. The synthetic part has different subsets that are either exact synthetic copies of the real data or are copies with certain aspects randomised. This allows us to analyse what types of variation are good for synthetic training data and which aspects should be modelled to closely match the target data. Furthermore, we investigate what types of training techniques are beneficial towards generalisation to real data, and how to use them. Additionally, we analyse how real images can be leveraged when training on synthetic images. All these experiments are validated on real data and benchmarked to models trained on real data. The results offer a number of interesting takeaways that can serve as basic guidelines for using synthetic data for object detection. Code to reproduce results is available at //github.com/EDM-Research/DIMO_ObjectDetection.
Pre-trained code generation models (PCGMs) have been widely applied in neural code generation which can generate executable code from functional descriptions in natural languages, possibly together with signatures. Despite substantial performance improvement of PCGMs, the role of method names in neural code generation has not been thoroughly investigated. In this paper, we study and demonstrate the potential of benefiting from method names to enhance the performance of PCGMs, from a model robustness perspective. Specifically, we propose a novel approach, named RADAR (neuRAl coDe generAtor Robustifier). RADAR consists of two components: RADAR-Attack and RADAR-Defense. The former attacks a PCGM by generating adversarial method names as part of the input, which are semantic and visual similar to the original input, but may trick the PCGM to generate completely unrelated code snippets. As a countermeasure to such attacks, RADAR-Defense synthesizes a new method name from the functional description and supplies it to the PCGM. Evaluation results show that RADAR-Attack can, e.g., reduce the CodeBLEU of generated code by 19.72% to 38.74% in three state-of-the-art PCGMs (i.e., CodeGPT, PLBART, and CodeT5). Moreover, RADAR-Defense is able to reinstate the performance of PCGMs with synthesized method names. These results highlight the importance of good method names in neural code generation and implicate the benefits of studying model robustness in software engineering.
Conditional independence and graphical models are well studied for probability distributions on product spaces. We propose a new notion of conditional independence for any measure $\Lambda$ on the punctured Euclidean space $\mathbb R^d\setminus \{0\}$ that explodes at the origin. The importance of such measures stems from their connection to infinitely divisible and max-infinitely divisible distributions, where they appear as L\'evy measures and exponent measures, respectively. We characterize independence and conditional independence for $\Lambda$ in various ways through kernels and factorization of a modified density, including a Hammersley-Clifford type theorem for undirected graphical models. As opposed to the classical conditional independence, our notion is intimately connected to the support of the measure $\Lambda$. Our general theory unifies and extends recent approaches to graphical modeling in the fields of extreme value analysis and L\'evy processes. Our results for the corresponding undirected and directed graphical models lay the foundation for new statistical methodology in these areas.
We introduce a Loss Discounting Framework for model and forecast combination which generalises and combines Bayesian model synthesis and generalized Bayes methodologies. We use a loss function to score the performance of different models and introduce a multilevel discounting scheme which allows a flexible specification of the dynamics of the model weights. This novel and simple model combination approach can be easily applied to large scale model averaging/selection, can handle unusual features such as sudden regime changes, and can be tailored to different forecasting problems. We compare our method to both established methodologies and state of the art methods for a number of macroeconomic forecasting examples. We find that the proposed method offers an attractive, computationally efficient alternative to the benchmark methodologies and often outperforms more complex techniques.
Over the last decade, a series of applied mathematics papers have explored a type of inverse problem--called by a variety of names including "inverse sensitivity", "pushforward based inference", "consistent Bayesian inference", or "data-consistent inversion"--wherein a solution is a probability density whose pushforward takes a given form. The formulation of such a stochastic inverse problem can be unexpected or confusing to those familiar with traditional Bayesian or otherwise statistical inference. To date, two classes of solutions have been proposed, and these have only been justified through applications of measure theory and its disintegration theorem. In this work we show that, under mild assumptions, the formulation of and solution to all stochastic inverse problems can be more clearly understood using basic probability theory: a stochastic inverse problem is simply a change-of-variables or approximation thereof. For the two existing classes of solutions, we derive the relationship to change(s)-of-variables and illustrate using analytic examples where none had previously existed. Our derivations use neither Bayes' theorem nor the disintegration theorem explicitly. Our final contribution is a careful comparison of changes-of-variables to more traditional statistical inference. While taking stochastic inverse problems at face value for the majority of the paper, our final comparative discussion gives a critique of the framework.
The dominating NLP paradigm of training a strong neural predictor to perform one task on a specific dataset has led to state-of-the-art performance in a variety of applications (eg. sentiment classification, span-prediction based question answering or machine translation). However, it builds upon the assumption that the data distribution is stationary, ie. that the data is sampled from a fixed distribution both at training and test time. This way of training is inconsistent with how we as humans are able to learn from and operate within a constantly changing stream of information. Moreover, it is ill-adapted to real-world use cases where the data distribution is expected to shift over the course of a model's lifetime. The first goal of this thesis is to characterize the different forms this shift can take in the context of natural language processing, and propose benchmarks and evaluation metrics to measure its effect on current deep learning architectures. We then proceed to take steps to mitigate the effect of distributional shift on NLP models. To this end, we develop methods based on parametric reformulations of the distributionally robust optimization framework. Empirically, we demonstrate that these approaches yield more robust models as demonstrated on a selection of realistic problems. In the third and final part of this thesis, we explore ways of efficiently adapting existing models to new domains or tasks. Our contribution to this topic takes inspiration from information geometry to derive a new gradient update rule which alleviate catastrophic forgetting issues during adaptation.
Modern neural network training relies heavily on data augmentation for improved generalization. After the initial success of label-preserving augmentations, there has been a recent surge of interest in label-perturbing approaches, which combine features and labels across training samples to smooth the learned decision surface. In this paper, we propose a new augmentation method that leverages the first and second moments extracted and re-injected by feature normalization. We replace the moments of the learned features of one training image by those of another, and also interpolate the target labels. As our approach is fast, operates entirely in feature space, and mixes different signals than prior methods, one can effectively combine it with existing augmentation methods. We demonstrate its efficacy across benchmark data sets in computer vision, speech, and natural language processing, where it consistently improves the generalization performance of highly competitive baseline networks.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191