Despite recent advances in automated machine learning, model selection is still a complex and computationally intensive process. For Gaussian processes (GPs), selecting the kernel is a crucial task, often done manually by the expert. Additionally, evaluating the model selection criteria for Gaussian processes typically scales cubically in the sample size, rendering kernel search particularly computationally expensive. We propose a novel, efficient search method through a general, structured kernel space. Previous methods solved this task via Bayesian optimization and relied on measuring the distance between GP's directly in function space to construct a kernel-kernel. We present an alternative approach by defining a kernel-kernel over the symbolic representation of the statistical hypothesis that is associated with a kernel. We empirically show that this leads to a computationally more efficient way of searching through a discrete kernel space.
The Bayesian transformed Gaussian process (BTG) model, proposed by Kedem and Oliviera, is a fully Bayesian counterpart to the warped Gaussian process (WGP) and marginalizes out a joint prior over input warping and kernel hyperparameters. This fully Bayesian treatment of hyperparameters often provides more accurate regression estimates and superior uncertainty propagation, but is prohibitively expensive. The BTG posterior predictive distribution, itself estimated through high-dimensional integration, must be inverted in order to perform model prediction. To make the Bayesian approach practical and comparable in speed to maximum-likelihood estimation (MLE), we propose principled and fast techniques for computing with BTG. Our framework uses doubly sparse quadrature rules, tight quantile bounds, and rank-one matrix algebra to enable both fast model prediction and model selection. These scalable methods allow us to regress over higher-dimensional datasets and apply BTG with layered transformations that greatly improve its expressibility. We demonstrate that BTG achieves superior empirical performance over MLE-based models.
Graph-based anomaly detection finds numerous applications in the real-world. Thus, there exists extensive literature on the topic that has recently shifted toward deep detection models due to advances in deep learning and graph neural networks (GNNs). A vast majority of prior work focuses on detecting node/edge/subgraph anomalies within a single graph, with much less work on graph-level anomaly detection in a graph database. This work aims to fill two gaps in the literature: We (1) design GLAM, an end-to-end graph-level anomaly detection model based on GNNs, and (2) focus on unsupervised model selection, which is notoriously hard due to lack of any labels, yet especially critical for deep NN based models with a long list of hyper-parameters. Further, we propose a new pooling strategy for graph-level embedding, called MMD-pooling, that is geared toward detecting distribution anomalies which has not been considered before. Through extensive experiments on 15 real-world datasets, we show that (i) GLAM outperforms node-level and two-stage (i.e. not end-to-end) baselines, and (ii) model selection picks a significantly more effective model than expectation (i.e. average) -- without using any labels -- among candidates with otherwise large variation in performance.
Neural networks are known to exploit spurious artifacts (or shortcuts) that co-occur with a target label, exhibiting heuristic memorization. On the other hand, networks have been shown to memorize training examples, resulting in example-level memorization. These kinds of memorization impede generalization of networks beyond their training distributions. Detecting such memorization could be challenging, often requiring researchers to curate tailored test sets. In this work, we hypothesize -- and subsequently show -- that the diversity in the activation patterns of different neurons is reflective of model generalization and memorization. We quantify the diversity in the neural activations through information-theoretic measures and find support for our hypothesis on experiments spanning several natural language and vision tasks. Importantly, we discover that information organization points to the two forms of memorization, even for neural activations computed on unlabelled in-distribution examples. Lastly, we demonstrate the utility of our findings for the problem of model selection. The associated code and other resources for this work are available at https://rachitbansal.github.io/information-measures.
Offline reinforcement learning (RL) can be used to improve future performance by leveraging historical data. There exist many different algorithms for offline RL, and it is well recognized that these algorithms, and their hyperparameter settings, can lead to decision policies with substantially differing performance. This prompts the need for pipelines that allow practitioners to systematically perform algorithm-hyperparameter selection for their setting. Critically, in most real-world settings, this pipeline must only involve the use of historical data. Inspired by statistical model selection methods for supervised learning, we introduce a task- and method-agnostic pipeline for automatically training, comparing, selecting, and deploying the best policy when the provided dataset is limited in size. In particular, our work highlights the importance of performing multiple data splits to produce more reliable algorithm-hyperparameter selection. While this is a common approach in supervised learning, to our knowledge, this has not been discussed in detail in the offline RL setting. We show it can have substantial impacts when the dataset is small. Compared to alternate approaches, our proposed pipeline outputs higher-performing deployed policies from a broad range of offline policy learning algorithms and across various simulation domains in healthcare, education, and robotics. This work contributes toward the development of a general-purpose meta-algorithm for automatic algorithm-hyperparameter selection for offline RL.
Algorithms for text-generation in dialogue can be misguided. For example, in task-oriented settings, reinforcement learning that optimizes only task-success can lead to abysmal lexical diversity. We hypothesize this is due to poor theoretical understanding of the objectives in text-generation and their relation to the learning process (i.e., model training). To this end, we propose a new theoretical framework for learning to generate text in dialogue. Compared to existing theories of learning, our framework allows for analysis of the multi-faceted goals inherent to text-generation. We use our framework to develop theoretical guarantees for learners that adapt to unseen data. As an example, we apply our theory to study data-shift within a cooperative learning algorithm proposed for the GuessWhat?! visual dialogue game. From this insight, we propose a new algorithm, and empirically, we demonstrate our proposal improves both task-success and human-likeness of the generated text. Finally, we show statistics from our theory are empirically predictive of multiple qualities of the generated dialogue, suggesting our theory is useful for model-selection when human evaluations are not available.
The calibration of deep learning-based perception models plays a crucial role in their reliability. Our work focuses on a class-wise evaluation of several model's confidence performance for LiDAR-based semantic segmentation with the aim of providing insights into the calibration of underrepresented classes. Those classes often include VRUs and are thus of particular interest for safety reasons. With the help of a metric based on sparsification curves we compare the calibration abilities of three semantic segmentation models with different architectural concepts, each in a in deterministic and a probabilistic version. By identifying and describing the dependency between the predictive performance of a class and the respective calibration quality we aim to facilitate the model selection and refinement for safety-critical applications.
Consider making a prediction over new test data without any opportunity to learn from a training set of labelled data - instead given access to a set of expert models and their predictions alongside some limited information about the dataset used to train them. In scenarios from finance to the medical sciences, and even consumer practice, stakeholders have developed models on private data they either cannot, or do not want to, share. Given the value and legislation surrounding personal information, it is not surprising that only the models, and not the data, will be released - the pertinent question becoming: how best to use these models? Previous work has focused on global model selection or ensembling, with the result of a single final model across the feature space. Machine learning models perform notoriously poorly on data outside their training domain however, and so we argue that when ensembling models the weightings for individual instances must reflect their respective domains - in other words models that are more likely to have seen information on that instance should have more attention paid to them. We introduce a method for such an instance-wise ensembling of models, including a novel representation learning step for handling sparse high-dimensional domains. Finally, we demonstrate the need and generalisability of our method on classical machine learning tasks as well as highlighting a real world use case in the pharmacological setting of vancomycin precision dosing.
When analyzing a dataset, it can be useful to assess how smooth the decision boundaries need to be for a model to better fit the data. This paper addresses this question by proposing the quantification of how much should the 'rigid' decision boundaries, produced by an algorithm that naturally finds such solutions, be relaxed to obtain a performance improvement. The approach we propose starts with the rigid decision boundaries of a seed Decision Tree (seed DT), which is used to initialize a Neural DT (NDT). The initial boundaries are challenged by relaxing them progressively through training the NDT. During this process, we measure the NDT's performance and decision agreement to its seed DT. We show how these two measures can help the user in figuring out how expressive his model should be, before exploring it further via model selection. The validity of our approach is demonstrated with experiments on simulated and benchmark datasets.
PARAGEN is a PyTorch-based NLP toolkit for further development on parallel generation. PARAGEN provides thirteen types of customizable plugins, helping users to experiment quickly with novel ideas across model architectures, optimization, and learning strategies. We implement various features, such as unlimited data loading and automatic model selection, to enhance its industrial usage. ParaGen is now deployed to support various research and industry applications at ByteDance. PARAGEN is available at https://github.com/bytedance/ParaGen.
Developing efficient time series anomaly detection techniques is important to maintain service quality and provide early alarms. Generative neural network methods are one class of the unsupervised approaches that are achieving increasing attention in recent years. In this paper, we have proposed an unsupervised Generative Adversarial Network (GAN)-based anomaly detection framework, DEGAN. It relies solely on normal time series data as input to train a well-configured discriminator (D) into a standalone anomaly predictor. In this framework, time series data is processed by the sliding window method. Expected normal patterns in data are leveraged to develop a generator (G) capable of generating normal data patterns. Normal data is also utilized in hyperparameter tuning and D model selection steps. Validated D models are then extracted and applied to evaluate unseen (testing) time series and identify patterns that have anomalous characteristics. Kernel density estimation (KDE) is applied to data points that are likely to be anomalous to generate probability density functions on the testing time series. The segments with the highest relative probabilities are detected as anomalies. To evaluate the performance, we tested on univariate acceleration time series for five miles of a Class I railroad track. We implemented the framework to detect the real anomalous observations identified by operators. The results show that leveraging the framework with a CNN D architecture results in average best recall and precision of 80% and 86%, respectively, which demonstrates that a well-trained standalone D model has the potential to be a reliable anomaly detector. Moreover, the influence of GAN hyperparameters, GAN architectures, sliding window sizes, clustering of time series, and model validation with labeled/unlabeled data were also investigated.
For more than 100 years, chemical, physical, and material scientists have proposed competing constitutive models to best characterize the behavior of natural and man-made materials in response to mechanical loading. Now, computer science offers a universal solution: Neural Networks. Neural Networks are powerful function approximators that can learn constitutive relations from large data without any knowledge of the underlying physics. However, classical Neural Networks ignore a century of research in constitutive modeling, violate thermodynamic considerations, and fail to predict the behavior outside the training regime. Here we design a new family of Constitutive Artificial Neural Networks that inherently satisfy common kinematic, thermodynamic, and physic constraints and, at the same time, constrain the design space of admissible functions to create robust approximators, even in the presence of sparse data. We revisit the non-linear field theories of mechanics and reverse-engineer the network input to account for material objectivity, symmetry, and incompressibility; the network output to enforce thermodynamic consistency; the activation functions to implement physically reasonable restrictions; and the network architecture to ensure polyconvexity. We demonstrate that this new class of models is a generalization of the classical neo Hooke, Blatz Ko, Mooney Rivlin, Yeoh, and Demiray models and that the network weights have a clear physical interpretation. When trained with classical benchmark data for rubber, our network autonomously selects the best constitutive model and learns its parameters. Our findings suggests that Constitutive Artificial Neural Networks have the potential to induce a paradigm shift in constitutive modeling, from user-defined model selection to automated model discovery. Our source code, data, and examples are available at https://github.com/LivingMatterLab/CANN.
Independent component analysis (ICA) is a blind source separation method for linear disentanglement of independent latent sources from observed data. We investigate the special setting of noisy linear ICA where the observations are split among different views, each receiving a mixture of shared and individual sources. We prove that the corresponding linear structure is identifiable, and the source distribution can be recovered. To computationally estimate the sources, we optimize a constrained form of the joint log-likelihood of the observed data among all views. We also show empirically that our objective recovers the sources also in the case when the measurements are corrupted by noise. Furthermore, we propose a model selection procedure for recovering the number of shared sources which we verify empirically. Finally, we apply the proposed model in a challenging real-life application, where the estimated shared sources from two large transcriptome datasets (observed data) provided by two different labs (two different views) lead to recovering (shared) sources utilized for finding a plausible representation of the underlying graph structure.
Unsupervised Domain Adaptation (UDA) aims at classifying unlabeled target images leveraging source labeled ones. In this work, we consider the Partial Domain Adaptation (PDA) variant, where we have extra source classes not present in the target domain. Most successful algorithms use model selection strategies that rely on target labels to find the best hyper-parameters and/or models along training. However, these strategies violate the main assumption in PDA: only unlabeled target domain samples are available. Moreover, there are also inconsistencies in the experimental settings - architecture, hyper-parameter tuning, number of runs - yielding unfair comparisons. The main goal of this work is to provide a realistic evaluation of PDA methods with the different model selection strategies under a consistent evaluation protocol. We evaluate 7 representative PDA algorithms on 2 different real-world datasets using 7 different model selection strategies. Our two main findings are: (i) without target labels for model selection, the accuracy of the methods decreases up to 30 percentage points; (ii) only one method and model selection pair performs well on both datasets. Experiments were performed with our PyTorch framework, BenchmarkPDA, which we open source.
Motivated by theoretical advancements in dimensionality reduction techniques we use a recent model, called Block Markov Chains, to conduct a practical study of clustering in real-world sequential data. Clustering algorithms for Block Markov Chains possess theoretical optimality guarantees and can be deployed in sparse data regimes. Despite these favorable theoretical properties, a thorough evaluation of these algorithms in realistic settings has been lacking. We address this issue and investigate the suitability of these clustering algorithms in exploratory data analysis of real-world sequential data. In particular, our sequential data is derived from human DNA, written text, animal movement data and financial markets. In order to evaluate the determined clusters, and the associated Block Markov Chain model, we further develop a set of evaluation tools. These tools include benchmarking, spectral noise analysis and statistical model selection tools. An efficient implementation of the clustering algorithm and the new evaluation tools is made available together with this paper. Practical challenges associated to real-world data are encountered and discussed. It is ultimately found that the Block Markov Chain model assumption, together with the tools developed here, can indeed produce meaningful insights in exploratory data analyses despite the complexity and sparsity of real-world data.
Probabilistic generative models provide a flexible and systematic framework for learning the underlying geometry of data. However, model selection in this setting is challenging, particularly when selecting for ill-defined qualities such as disentanglement or interpretability. In this work, we address this gap by introducing a method for ranking generative models based on the training dynamics exhibited during learning. Inspired by recent theoretical characterizations of disentanglement, our method does not require supervision of the underlying latent factors. We evaluate our approach by demonstrating the need for disentanglement metrics which do not require labels\textemdash the underlying generative factors. We additionally demonstrate that our approach correlates with baseline supervised methods for evaluating disentanglement. Finally, we show that our method can be used as an unsupervised indicator for downstream performance on reinforcement learning and fairness-classification problems.
A multitude of work has shown that machine learning-based medical diagnosis systems can be biased against certain subgroups of people. This has motivated a growing number of bias mitigation algorithms that aim to address fairness issues in machine learning. However, it is difficult to compare their effectiveness in medical imaging for two reasons. First, there is little consensus on the criteria to assess fairness. Second, existing bias mitigation algorithms are developed under different settings, e.g., datasets, model selection strategies, backbones, and fairness metrics, making a direct comparison and evaluation based on existing results impossible. In this work, we introduce MEDFAIR, a framework to benchmark the fairness of machine learning models for medical imaging. MEDFAIR covers eleven algorithms from various categories, nine datasets from different imaging modalities, and three model selection criteria. Through extensive experiments, we find that the under-studied issue of model selection criterion can have a significant impact on fairness outcomes; while in contrast, state-of-the-art bias mitigation algorithms do not significantly improve fairness outcomes over empirical risk minimization (ERM) in both in-distribution and out-of-distribution settings. We evaluate fairness from various perspectives and make recommendations for different medical application scenarios that require different ethical principles. Our framework provides a reproducible and easy-to-use entry point for the development and evaluation of future bias mitigation algorithms in deep learning. Code is available at https://github.com/ys-zong/MEDFAIR.
Federated Learning (FL) framework enables training over distributed datasets while keeping the data local. However, it is difficult to customize a model fitting for all unknown local data. A pre-determined model is most likely to lead to slow convergence or low accuracy, especially when the distributed data is non-i.i.d.. To resolve the issue, we propose a model searching method in the federated learning scenario, and the method automatically searches a model structure fitting for the unseen local data. We novelly design a reinforcement learning-based framework that samples and distributes sub-models to the participants and updates its model selection policy by maximizing the reward. In practice, the model search algorithm takes a long time to converge, and hence we adaptively assign sub-models to participants according to the transmission condition. We further propose delay-compensated synchronization to mitigate loss over late updates to facilitate convergence. Extensive experiments show that our federated model search algorithm produces highly accurate models efficiently, particularly on non-i.i.d. data.
A prevalent feature of high-dimensional data is the dependence among covariates, and model selection is known to be challenging when covariates are highly correlated. To perform model selection for the high-dimensional Cox proportional hazards model in presence of correlated covariates with factor structure, we propose a new model, Factor-Augmented Regularized Model for Hazard Regression (FarmHazard), which builds upon latent factors that drive covariate dependence and extends Cox's model. This new model generates procedures that operate in two steps by learning factors and idiosyncratic components from high-dimensional covariate vectors and then using them as new predictors. Cox's model is a widely used semi-parametric model for survival analysis, where censored data and time-dependent covariates bring additional technical challenges. We prove model selection consistency and estimation consistency under mild conditions. We also develop a factor-augmented variable screening procedure to deal with strong correlations in ultra-high dimensional problems. Extensive simulations and real data experiments demonstrate that our procedures enjoy good performance and achieve better results on model selection, out-of-sample C-index and screening than alternative methods.
Anomaly detection in time-series has a wide range of practical applications. While numerous anomaly detection methods have been proposed in the literature, a recent survey concluded that no single method is the most accurate across various datasets. To make matters worse, anomaly labels are scarce and rarely available in practice. The practical problem of selecting the most accurate model for a given dataset without labels has received little attention in the literature. This paper answers this question i.e. Given an unlabeled dataset and a set of candidate anomaly detectors, how can we select the most accurate model? To this end, we identify three classes of surrogate (unsupervised) metrics, namely, prediction error, model centrality, and performance on injected synthetic anomalies, and show that some metrics are highly correlated with standard supervised anomaly detection performance metrics such as the $F_1$ score, but to varying degrees. We formulate metric combination with multiple imperfect surrogate metrics as a robust rank aggregation problem. We then provide theoretical justification behind the proposed approach. Large-scale experiments on multiple real-world datasets demonstrate that our proposed unsupervised approach is as effective as selecting the most accurate model based on partially labeled data.