Consider the spiked Wigner model \[ X = \sum_{i = 1}^k \lambda_i u_i u_i^\top + \sigma G, \] where $G$ is an $N \times N$ GOE random matrix, and the eigenvalues $\lambda_i$ are all spiked, i.e. above the Baik-Ben Arous-P\'ech\'e (BBP) threshold $\sigma$. We consider AIC-type model selection criteria of the form \[ -2 \, (\text{maximised log-likelihood}) + \gamma \, (\text{number of parameters}) \] for estimating the number $k$ of spikes. For $\gamma > 2$, the above criterion is strongly consistent provided $\lambda_k > \lambda_{\gamma}$, where $\lambda_{\gamma}$ is a threshold strictly above the BBP threshold, whereas for $\gamma < 2$, it almost surely overestimates $k$. Although AIC (which corresponds to $\gamma = 2$) is not strongly consistent, we show that taking $\gamma = 2 + \delta_N$, where $\delta_N \to 0$ and $\delta_N \gg N^{-2/3}$, results in a weakly consistent estimator of $k$. We further show that a soft minimiser of AIC, where one chooses the least complex model whose AIC score is close to the minimum AIC score, is strongly consistent. Based on a spiked (generalised) Wigner representation, we also develop similar model selection criteria for consistently estimating the number of communities in a balanced stochastic block model under some sparsity restrictions.
Outlier detection (OD) finds many applications with a rich literature of numerous techniques. Deep neural network based OD (DOD) has seen a recent surge of attention thanks to the many advances in deep learning. In this paper, we consider a critical-yet-understudied challenge with unsupervised DOD, that is, effective hyperparameter (HP) tuning/model selection. While several prior work report the sensitivity of OD models to HPs, it becomes ever so critical for the modern DOD models that exhibit a long list of HPs. We introduce HYPER for tuning DOD models, tackling two fundamental challenges: (1) validation without supervision (due to lack of labeled anomalies), and (2) efficient search of the HP/model space (due to exponential growth in the number of HPs). A key idea is to design and train a novel hypernetwork (HN) that maps HPs onto optimal weights of the main DOD model. In turn, HYPER capitalizes on a single HN that can dynamically generate weights for many DOD models (corresponding to varying HPs), which offers significant speed-up. In addition, it employs meta-learning on historical OD tasks with labels to train a proxy validation function, likewise trained with our proposed HN efficiently. Extensive experiments on 35 OD tasks show that HYPER achieves high performance against 8 baselines with significant efficiency gains.
As causal ground truth is incredibly rare, causal discovery algorithms are commonly only evaluated on simulated data. This is concerning, given that simulations reflect preconceptions about generating processes regarding noise distributions, model classes, and more. In this work, we propose a novel method for falsifying the output of a causal discovery algorithm in the absence of ground truth. Our key insight is that while statistical learning seeks stability across subsets of data points, causal learning should seek stability across subsets of variables. Motivated by this insight, our method relies on a notion of compatibility between causal graphs learned on different subsets of variables. We prove that detecting incompatibilities can falsify wrongly inferred causal relations due to violation of assumptions or errors from finite sample effects. Although passing such compatibility tests is only a necessary criterion for good performance, we argue that it provides strong evidence for the causal models whenever compatibility entails strong implications for the joint distribution. We also demonstrate experimentally that detection of incompatibilities can aid in causal model selection.
This short paper presents a preliminary analysis of three popular Visual Question Answering (VQA) models, namely ViLBERT, ViLT, and LXMERT, in the context of answering questions relating to driving scenarios. The performance of these models is evaluated by comparing the similarity of responses to reference answers provided by computer vision experts. Model selection is predicated on the analysis of transformer utilization in multimodal architectures. The results indicate that models incorporating cross-modal attention and late fusion techniques exhibit promising potential for generating improved answers within a driving perspective. This initial analysis serves as a launchpad for a forthcoming comprehensive comparative study involving nine VQA models and sets the scene for further investigations into the effectiveness of VQA model queries in self-driving scenarios. Supplementary material is available at https://github.com/KaavyaRekanar/Towards-a-performance-analysis-on-pre-trained-VQA-models-for-autonomous-driving.
The problem of model selection is considered for the setting of interpolating estimators, where the number of model parameters exceeds the size of the dataset. Classical information criteria typically consider the large-data limit, penalizing model size. However, these criteria are not appropriate in modern settings where overparameterized models tend to perform well. For any overparameterized model, we show that there exists a dual underparameterized model that possesses the same marginal likelihood, thus establishing a form of Bayesian duality. This enables more classical methods to be used in the overparameterized setting, revealing the Interpolating Information Criterion, a measure of model quality that naturally incorporates the choice of prior into the model selection. Our new information criterion accounts for prior misspecification, geometric and spectral properties of the model, and is numerically consistent with known empirical and theoretical behavior in this regime.
Statistical inference of the high-dimensional regression coefficients is challenging because the uncertainty introduced by the model selection procedure is hard to account for. A critical question remains unsettled; that is, is it possible and how to embed the inference of the model into the simultaneous inference of the coefficients? To this end, we propose a notion of simultaneous confidence intervals called the sparsified simultaneous confidence intervals. Our intervals are sparse in the sense that some of the intervals' upper and lower bounds are shrunken to zero (i.e., $[0,0]$), indicating the unimportance of the corresponding covariates. These covariates should be excluded from the final model. The rest of the intervals, either containing zero (e.g., $[-1,1]$ or $[0,1]$) or not containing zero (e.g., $[2,3]$), indicate the plausible and significant covariates, respectively. The proposed method can be coupled with various selection procedures, making it ideal for comparing their uncertainty. For the proposed method, we establish desirable asymptotic properties, develop intuitive graphical tools for visualization, and justify its superior performance through simulation and real data analysis.
In out-of-distribution (OOD) generalization tasks, fine-tuning pre-trained models has become a prevalent strategy. Different from most prior work that has focused on advancing learning algorithms, we systematically examined how pre-trained model size, pre-training dataset size, and training strategies impact generalization and uncertainty calibration on downstream tasks. We evaluated 100 models across diverse pre-trained model sizes, \update{five} pre-training datasets, and five data augmentations through extensive experiments on four distribution shift datasets totaling over 120,000 GPU hours. Our results demonstrate the significant impact of pre-trained model selection, with optimal choices substantially improving OOD accuracy over algorithm improvement alone. We find larger models and bigger pre-training data improve OOD performance and calibration, in contrast to some prior studies that found modern deep networks to calibrate worse than classical shallow models. Our work underscores the overlooked importance of pre-trained model selection for out-of-distribution generalization and calibration.
Most image retrieval research prioritizes improving predictive performance, often overlooking situations where the reliability of predictions is equally important. The gap between model performance and reliability requirements highlights the need for a systematic approach to analyze and address the risks associated with image retrieval. Uncertainty quantification technique can be applied to mitigate this issue by assessing uncertainty for retrieval sets, but it provides only a heuristic estimate of uncertainty rather than a guarantee. To address these limitations, we present Risk Controlled Image Retrieval (RCIR), which generates retrieval sets with coverage guarantee, i.e., retrieval sets that are guaranteed to contain the true nearest neighbors with a predefined probability. RCIR can be easily integrated with existing uncertainty-aware image retrieval systems, agnostic to data distribution and model selection. To the best of our knowledge, this is the first work that provides coverage guarantees to image retrieval. The validity and efficiency of RCIR are demonstrated on four real-world datasets: CAR-196, CUB-200, Pittsburgh, and ChestX-Det.
Current automated machine learning (ML) tools are model-centric, focusing on model selection and parameter optimization. However, the majority of the time in data analysis is devoted to data cleaning and wrangling, for which limited tools are available. Here we present DataAssist, an automated data preparation and cleaning platform that enhances dataset quality using ML-informed methods. We show that DataAssist provides a pipeline for exploratory data analysis and data cleaning, including generating visualization for user-selected variables, unifying data annotation, suggesting anomaly removal, and preprocessing data. The exported dataset can be readily integrated with other autoML tools or user-specified model for downstream analysis. Our data-centric tool is applicable to a variety of fields, including economics, business, and forecasting applications saving over 50% time of the time spent on data cleansing and preparation.
Self-supervised learning (SSL) has proven effective in solving various problems by generating internal supervisory signals. Unsupervised anomaly detection, which faces the high cost of obtaining true labels, is an area that can greatly benefit from SSL. However, recent literature suggests that tuning the hyperparameters (HP) of data augmentation functions is crucial to the success of SSL-based anomaly detection (SSAD), yet a systematic method for doing so remains unknown. In this work, we propose DSV (Discordance and Separability Validation), an unsupervised validation loss to select high-performing detection models with effective augmentation HPs. DSV captures the alignment between an augmentation function and the anomaly-generating mechanism with surrogate losses, which approximate the discordance and separability of test data, respectively. As a result, the evaluation via DSV leads to selecting an effective SSAD model exhibiting better alignment, which results in high detection accuracy. We theoretically derive the degree of approximation conducted by the surrogate losses and empirically show that DSV outperforms a wide range of baselines on 21 real-world tasks.
The vulnerability of Deep Neural Networks (DNNs) to adversarial attacks poses a significant challenge to their deployment in safety-critical applications. While extensive research has addressed various attack scenarios, the no-box attack setting where adversaries have no prior knowledge, including access to training data of the target model, remains relatively underexplored despite its practical relevance. This work presents a systematic investigation into leveraging large-scale Vision-Language Models (VLMs), particularly CLIP, as surrogate models for executing no-box attacks. Our theoretical and empirical analyses reveal a key limitation in the execution of no-box attacks stemming from insufficient discriminative capabilities for direct application of vanilla CLIP as a surrogate model. To address this limitation, we propose MF-CLIP: a novel framework that enhances CLIP's effectiveness as a surrogate model through margin-aware feature space optimization. Comprehensive evaluations across diverse architectures and datasets demonstrate that MF-CLIP substantially advances the state-of-the-art in no-box attacks, surpassing existing baselines by 15.23% on standard models and achieving a 9.52% improvement on adversarially trained models. Our code will be made publicly available to facilitate reproducibility and future research in this direction.
The number of modes in a probability density function is representative of the complexity of a model and can also be viewed as the number of subpopulations. Despite its relevance, there has been limited research in this area. A novel approach to estimating the number of modes in the univariate setting is presented, focusing on prediction accuracy and inspired by some overlooked aspects of the problem: the need for structure in the solutions, the subjective and uncertain nature of modes, and the convenience of a holistic view that blends local and global density properties. The technique combines flexible kernel estimators and parsimonious compositional splines in the Bayesian inference paradigm, providing soft solutions and incorporating expert judgment. The procedure includes feature exploration, model selection, and mode testing, illustrated in a sports analytics case study showcasing multiple companion visualisation tools. A thorough simulation study also demonstrates that traditional modality-driven approaches paradoxically struggle to provide accurate results. In this context, the new method emerges as a top-tier alternative, offering innovative solutions for analysts.
Recent progress in research on Deep Graph Networks (DGNs) has led to a maturation of the domain of learning on graphs. Despite the growth of this research field, there are still important challenges that are yet unsolved. Specifically, there is an urge of making DGNs suitable for predictive tasks on realworld systems of interconnected entities, which evolve over time. With the aim of fostering research in the domain of dynamic graphs, at first, we survey recent advantages in learning both temporal and spatial information, providing a comprehensive overview of the current state-of-the-art in the domain of representation learning for dynamic graphs. Secondly, we conduct a fair performance comparison among the most popular proposed approaches on node and edge-level tasks, leveraging rigorous model selection and assessment for all the methods, thus establishing a sound baseline for evaluating new architectures and approaches
The Laplace approximation provides a closed-form model selection objective for neural networks (NN). Online variants, which optimise NN parameters jointly with hyperparameters, like weight decay strength, have seen renewed interest in the Bayesian deep learning community. However, these methods violate Laplace's method's critical assumption that the approximation is performed around a mode of the loss, calling into question their soundness. This work re-derives online Laplace methods, showing them to target a variational bound on a mode-corrected variant of the Laplace evidence which does not make stationarity assumptions. Online Laplace and its mode-corrected counterpart share stationary points where 1. the NN parameters are a maximum a posteriori, satisfying the Laplace method's assumption, and 2. the hyperparameters maximise the Laplace evidence, motivating online methods. We demonstrate that these optima are roughly attained in practise by online algorithms using full-batch gradient descent on UCI regression datasets. The optimised hyperparameters prevent overfitting and outperform validation-based early stopping.
In the context of the rapid development of large language models, we have meticulously trained and introduced the GujiBERT and GujiGPT language models, which are foundational models specifically designed for intelligent information processing of ancient texts. These models have been trained on an extensive dataset that encompasses both simplified and traditional Chinese characters, allowing them to effectively handle various natural language processing tasks related to ancient books, including but not limited to automatic sentence segmentation, punctuation, word segmentation, part-of-speech tagging, entity recognition, and automatic translation. Notably, these models have exhibited exceptional performance across a range of validation tasks using publicly available datasets. Our research findings highlight the efficacy of employing self-supervised methods to further train the models using classical text corpora, thus enhancing their capability to tackle downstream tasks. Moreover, it is worth emphasizing that the choice of font, the scale of the corpus, and the initial model selection all exert significant influence over the ultimate experimental outcomes. To cater to the diverse text processing preferences of researchers in digital humanities and linguistics, we have developed three distinct categories comprising a total of nine model variations. We believe that by sharing these foundational language models specialized in the domain of ancient texts, we can facilitate the intelligent processing and scholarly exploration of ancient literary works and, consequently, contribute to the global dissemination of China's rich and esteemed traditional culture in this new era.
Numerous benchmarks for Few-Shot Learning have been proposed in the last decade. However all of these benchmarks focus on performance averaged over many tasks, and the question of how to reliably evaluate and tune models trained for individual tasks in this regime has not been addressed. This paper presents the first investigation into task-level evaluation -- a fundamental step when deploying a model. We measure the accuracy of performance estimators in the few-shot setting, consider strategies for model selection, and examine the reasons for the failure of evaluators usually thought of as being robust. We conclude that cross-validation with a low number of folds is the best choice for directly estimating the performance of a model, whereas using bootstrapping or cross validation with a large number of folds is better for model selection purposes. Overall, we find that existing benchmarks for few-shot learning are not designed in such a way that one can get a reliable picture of how effectively methods can be used on individual tasks.
The problem of detecting and quantifying the presence of symmetries in datasets is useful for model selection, generative modeling, and data analysis, amongst others. While existing methods for hard-coding transformations in neural networks require prior knowledge of the symmetries of the task at hand, this work focuses on discovering and characterizing unknown symmetries present in the dataset, namely, Lie group symmetry transformations beyond the traditional ones usually considered in the field (rotation, scaling, and translation). Specifically, we consider a scenario in which a dataset has been transformed by a one-parameter subgroup of transformations with different parameter values for each data point. Our goal is to characterize the transformation group and the distribution of the parameter values. The results showcase the effectiveness of the approach in both these settings.
Proximal nested sampling was introduced recently to open up Bayesian model selection for high-dimensional problems such as computational imaging. The framework is suitable for models with a log-convex likelihood, which are ubiquitous in the imaging sciences. The purpose of this article is two-fold. First, we review proximal nested sampling in a pedagogical manner in an attempt to elucidate the framework for physical scientists. Second, we show how proximal nested sampling can be extended in an empirical Bayes setting to support data-driven priors, such as deep neural networks learned from training data.
Computing the marginal likelihood (also called the Bayesian model evidence) is an important task in Bayesian model selection, providing a principled quantitative way to compare models. The learned harmonic mean estimator solves the exploding variance problem of the original harmonic mean estimation of the marginal likelihood. The learned harmonic mean estimator learns an importance sampling target distribution that approximates the optimal distribution. While the approximation need not be highly accurate, it is critical that the probability mass of the learned distribution is contained within the posterior in order to avoid the exploding variance problem. In previous work a bespoke optimization problem is introduced when training models in order to ensure this property is satisfied. In the current article we introduce the use of normalizing flows to represent the importance sampling target distribution. A flow-based model is trained on samples from the posterior by maximum likelihood estimation. Then, the probability density of the flow is concentrated by lowering the variance of the base distribution, i.e. by lowering its "temperature", ensuring its probability mass is contained within the posterior. This approach avoids the need for a bespoke optimisation problem and careful fine tuning of parameters, resulting in a more robust method. Moreover, the use of normalizing flows has the potential to scale to high dimensional settings. We present preliminary experiments demonstrating the effectiveness of the use of flows for the learned harmonic mean estimator. The harmonic code implementing the learned harmonic mean, which is publicly available, has been updated to now support normalizing flows.
Many state-of-the-art automated machine learning (AutoML) systems use greedy ensemble selection (GES) by Caruana et al. (2004) to ensemble models found during model selection post hoc. Thereby, boosting predictive performance and likely following Auto-Sklearn 1's insight that alternatives, like stacking or gradient-free numerical optimization, overfit. Overfitting in Auto-Sklearn 1 is much more likely than in other AutoML systems because it uses only low-quality validation data for post hoc ensembling. Therefore, we were motivated to analyze whether Auto-Sklearn 1's insight holds true for systems with higher-quality validation data. Consequently, we compared the performance of covariance matrix adaptation evolution strategy (CMA-ES), state-of-the-art gradient-free numerical optimization, to GES on the 71 classification datasets from the AutoML benchmark for AutoGluon. We found that Auto-Sklearn's insight depends on the chosen metric. For the metric ROC AUC, CMA-ES overfits drastically and is outperformed by GES -- statistically significantly for multi-class classification. For the metric balanced accuracy, CMA-ES does not overfit and outperforms GES significantly. Motivated by the successful application of CMA-ES for balanced accuracy, we explored methods to stop CMA-ES from overfitting for ROC AUC. We propose a method to normalize the weights produced by CMA-ES, inspired by GES, that avoids overfitting for CMA-ES and makes CMA-ES perform better than or similar to GES for ROC AUC.