Large-scale vision-language models (VLMs) like CLIP successfully find correspondences between images and text. Through the standard deterministic mapping process, an image or a text sample is mapped to a single vector in the embedding space. This is problematic: as multiple samples (images or text) can abstract the same concept in the physical world, deterministic embeddings do not reflect the inherent ambiguity in the embedding space. We propose ProbVLM, a probabilistic adapter that estimates probability distributions for the embeddings of pre-trained VLMs via inter/intra-modal alignment in a post-hoc manner without needing large-scale datasets or computing. On four challenging datasets, i.e., COCO, Flickr, CUB, and Oxford-flowers, we estimate the multi-modal embedding uncertainties for two VLMs, i.e., CLIP and BLIP, quantify the calibration of embedding uncertainties in retrieval tasks and show that ProbVLM outperforms other methods. Furthermore, we propose active learning and model selection as two real-world downstream tasks for VLMs and show that the estimated uncertainty aids both tasks. Lastly, we present a novel technique for visualizing the embedding distributions using a large-scale pre-trained latent diffusion model. Code is available at https://github.com/ExplainableML/ProbVLM.
Inferring the mechanisms underlying physiological and pathological processes in the brain from recorded electrical activity is challenging. Bayesian model selection and dynamic causal modelling aim to identify likely biophysical models to explain data and to fit the model parameters. Here, we use data generated by simulations to investigate the effectiveness of Bayesian model selection and dynamic causal modelling when applied at steady state in the frequency domain to identify and fit Jansen-Rit models. We first investigate the impact of the necessary assumption of linearity on the dynamics of the Jansen-Rit model. We then apply dynamic causal modelling and Bayesian model selection to data generated from simulations of linear neural mass models, non-linear neural mass models, and networks of discrete spiking neurons. Action potentials are a characteristic feature of neuronal dynamics but have not previously been explicitly included in simulations used to test Bayesian model selection or dynamic causal modelling. We find that the assumption of linearity abolishes the qualitative transitions seen as a function of the connectivity parameter in the original Jansen-Rit model. As with previous work, we find that the recovery procedures are effective when applied to data from linear Jansen-Rit neural mass models, however, when applying them to non-linear neural mass models and networks of discrete spiking neurons we find that their effectiveness is significantly reduced, suggesting caution is required when applying these methods.
The number of end-to-end speech recognition models grows every year. These models are often adapted to new domains or languages resulting in a proliferation of expert systems that achieve great results on target data, while generally showing inferior performance outside of their domain of expertise. We explore combination of such experts via confidence-based ensembles: ensembles of models where only the output of the most-confident model is used. We assume that models' target data is not available except for a small validation set. We demonstrate effectiveness of our approach with two applications. First, we show that a confidence-based ensemble of 5 monolingual models outperforms a system where model selection is performed via a dedicated language identification block. Second, we demonstrate that it is possible to combine base and adapted models to achieve strong results on both original and target data. We validate all our results on multiple datasets and model architectures.
Predictive pattern mining is an approach used to construct prediction models when the input is represented by structured data, such as sets, graphs, and sequences. The main idea behind predictive pattern mining is to build a prediction model by considering substructures, such as subsets, subgraphs, and subsequences (referred to as patterns), present in the structured data as features of the model. The primary challenge in predictive pattern mining lies in the exponential growth of the number of patterns with the complexity of the structured data. In this study, we propose the Safe Pattern Pruning (SPP) method to address the explosion of pattern numbers in predictive pattern mining. We also discuss how it can be effectively employed throughout the entire model building process in practical data analysis. To demonstrate the effectiveness of the proposed method, we conduct numerical experiments on regression and classification problems involving sets, graphs, and sequences.
In this study, we have developed an incremental machine learning (ML) method that efficiently obtains the optimal model when a small number of instances or features are added or removed. This problem holds practical importance in model selection, such as cross-validation (CV) and feature selection. Among the class of ML methods known as linear estimators, there exists an efficient model update framework called the low-rank update that can effectively handle changes in a small number of rows and columns within the data matrix. However, for ML methods beyond linear estimators, there is currently no comprehensive framework available to obtain knowledge about the updated solution within a specific computational complexity. In light of this, our study introduces a method called the Generalized Low-Rank Update (GLRU) which extends the low-rank update framework of linear estimators to ML methods formulated as a certain class of regularized empirical risk minimization, including commonly used methods such as SVM and logistic regression. The proposed GLRU method not only expands the range of its applicability but also provides information about the updated solutions with a computational complexity proportional to the amount of dataset changes. To demonstrate the effectiveness of the GLRU method, we conduct experiments showcasing its efficiency in performing cross-validation and feature selection compared to other baseline methods.
Deep neural networks often exploit *spurious* features that are present in the majority of examples within a class during training. This leads to *poor worst-group test accuracy*, i.e., poor accuracy for minority groups that lack these spurious features. Despite the growing body of recent efforts to address spurious correlations (SC), several challenging settings remain unexplored.In this work, we propose studying methods to mitigate SC in settings with: 1) spurious features that are learned more slowly, 2) a larger number of classes, and 3) a larger number of groups. We introduce two new datasets, Animals and SUN, to facilitate this study and conduct a systematic benchmarking of 8 state-of-the-art (SOTA) methods across a total of 5 vision datasets, training over 5,000 models. Through this, we highlight how existing group inference methods struggle in the presence of spurious features that are learned later in training. Additionally, we demonstrate how all existing methods struggle in settings with more groups and/or classes. Finally, we show the importance of careful model selection (hyperparameter tuning) in extracting optimal performance, especially in the more challenging settings we introduced, and propose more cost-efficient strategies for model selection. Overall, through extensive and systematic experiments, this work uncovers a suite of new challenges and opportunities for improving worst-group generalization in the presence of spurious features. Our datasets, methods and scripts available at https://github.com/BigML-CS-UCLA/SpuCo.
In data-driven optimization, the sample performance of the obtained decision typically incurs an optimistic bias against the true performance, a phenomenon commonly known as the Optimizer's Curse and intimately related to overfitting in machine learning. Common techniques to correct this bias, such as cross-validation, require repeatedly solving additional optimization problems and are therefore computationally expensive. We develop a general bias correction approach, building on what we call Optimizer's Information Criterion (OIC), that directly approximates the first-order bias and does not require solving any additional optimization problems. Our OIC generalizes the celebrated Akaike Information Criterion to evaluate the objective performance in data-driven optimization, which crucially involves not only model fitting but also its interplay with the downstream optimization. As such it can be used for decision selection instead of only model selection. We apply our approach to a range of data-driven optimization formulations comprising empirical and parametric models, their regularized counterparts, and furthermore contextual optimization. Finally, we provide numerical validation on the superior performance of our approach under synthetic and real-world datasets.
The preservation of soil health is a critical challenge in the 21st century due to its significant impact on agriculture, human health, and biodiversity. We provide the first deep investigation of the predictive potential of machine learning models to understand the connections between soil and biological phenotypes. We investigate an integrative framework performing accurate machine learning-based prediction of plant phenotypes from biological, chemical, and physical properties of the soil via two models: random forest and Bayesian neural network. We show that prediction is improved when incorporating environmental features like soil physicochemical properties and microbial population density into the models, in addition to the microbiome information. Exploring various data preprocessing strategies confirms the significant impact of human decisions on predictive performance. We show that the naive total sum scaling normalization that is commonly used in microbiome research is not the optimal strategy to maximize predictive power. Also, we find that accurately defined labels are more important than normalization, taxonomic level or model characteristics. In cases where humans are unable to classify samples accurately, machine learning model performance is limited. Lastly, we provide domain scientists via a full model selection decision tree to identify the human choices that optimize model prediction power. Our work is accompanied by open source reproducible scripts (https://github.com/solislemuslab/soil-microbiome-nn) for maximum outreach among the microbiome research community.
In LiDAR-based 3D detection, history point clouds contain rich temporal information helpful for future prediction. In the same way, history detections should contribute to future detections. In this paper, we propose a detection enhancement method, namely FrameFusion, which improves 3D object detection results by fusing history frames. In FrameFusion, we ''forward'' history frames to the current frame and apply weighted Non-Maximum-Suppression on dense bounding boxes to obtain a fused frame with merged boxes. To ''forward'' frames, we use vehicle motion models to estimate the future pose of the bounding boxes. However, the commonly used constant velocity model fails naturally on turning vehicles, so we explore two vehicle motion models to address this issue. On Waymo Open Dataset, our FrameFusion method consistently improves the performance of various 3D detectors by about $2$ vehicle level 2 APH with negligible latency and slightly enhances the performance of the temporal fusion method MPPNet. We also conduct extensive experiments on motion model selection.
Effective machine learning models learn both robust features that directly determine the outcome of interest (e.g., an object with wheels is more likely to be a car), and shortcut features (e.g., an object on a road is more likely to be a car). The latter can be a source of error under distributional shift, when the correlations change at test-time. The prevailing sentiment in the robustness literature is to avoid such correlative shortcut features and learn robust predictors. However, while robust predictors perform better on worst-case distributional shifts, they often sacrifice accuracy on majority subpopulations. In this paper, we argue that shortcut features should not be entirely discarded. Instead, if we can identify the subpopulation to which an input belongs, we can adaptively choose among models with different strengths to achieve high performance on both majority and minority subpopulations. We propose COnfidence-baSed MOdel Selection (CosMoS), where we observe that model confidence can effectively guide model selection. Notably, CosMoS does not require any target labels or group annotations, either of which may be difficult to obtain or unavailable. We evaluate CosMoS on four datasets with spurious correlations, each with multiple test sets with varying levels of data distribution shift. We find that CosMoS achieves 2-5% lower average regret across all subpopulations, compared to using only robust predictors or other model aggregation methods.
In the pursuit of further advancement in the field of target tracking, this paper explores the efficacy of a feedforward neural network in predicting drones tracks, aiming to eventually, compare the tracks created by the well-known Kalman filter and the ones created by our proposed neural network. The unique feature of our proposed neural network tracker is that it is using only a measurement model to estimate the next states of the track. Object model selection and linearization is one of the challenges that always face in the tracking process. The neural network uses a sliding window to incorporate the history of measurements when applying estimations of the track values. The testing results are comparable to the ones generated by the Kalman filter, especially for the cases where there is low measurement covariance. The complexity of linearization is avoided when using this proposed model.
Histograms are among the most popular methods used in exploratory analysis to summarize univariate distributions. In particular, irregular histograms are good non-parametric density estimators that require very few parameters: the number of bins with their lengths and frequencies. Many approaches have been proposed in the literature to infer these parameters, either assuming hypotheses about the underlying data distributions or exploiting a model selection approach. In this paper, we focus on the G-Enum histogram method, which exploits the Minimum Description Length (MDL) principle to build histograms without any user parameter and achieves state-of-the art performance w.r.t accuracy; parsimony and computation time. We investigate on the limits of this method in the case of outliers or heavy-tailed distributions. We suggest a two-level heuristic to deal with such cases. The first level exploits a logarithmic transformation of the data to split the data set into a list of data subsets with a controlled range of values. The second level builds a sub-histogram for each data subset and aggregates them to obtain a complete histogram. Extensive experiments show the benefits of the approach.
Double-descent refers to the unexpected drop in test loss of a learning algorithm beyond an interpolating threshold with over-parameterization, which is not predicted by information criteria in their classical forms due to the limitations in the standard asymptotic approach. We update these analyses using the information risk minimization framework and provide Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) for models learned by the Gibbs algorithm. Notably, the penalty terms for the Gibbs-based AIC and BIC correspond to specific information measures, i.e., symmetrized KL information and KL divergence. We extend this information-theoretic analysis to over-parameterized models by providing two different Gibbs-based BICs to compute the marginal likelihood of random feature models in the regime where the number of parameters $p$ and the number of samples $n$ tend to infinity, with $p/n$ fixed. Our experiments demonstrate that the Gibbs-based BIC can select the high-dimensional model and reveal the mismatch between marginal likelihood and population risk in the over-parameterized regime, providing new insights to understand double-descent.
Selecting hyperparameters in deep learning greatly impacts its effectiveness but requires manual effort and expertise. Recent works show that Bayesian model selection with Laplace approximations can allow to optimize such hyperparameters just like standard neural network parameters using gradients and on the training data. However, estimating a single hyperparameter gradient requires a pass through the entire dataset, limiting the scalability of such algorithms. In this work, we overcome this issue by introducing lower bounds to the linearized Laplace approximation of the marginal likelihood. In contrast to previous estimators, these bounds are amenable to stochastic-gradient-based optimization and allow to trade off estimation accuracy against computational complexity. We derive them using the function-space form of the linearized Laplace, which can be estimated using the neural tangent kernel. Experimentally, we show that the estimators can significantly accelerate gradient-based hyperparameter optimization.
Test-Time Adaptation (TTA) has recently emerged as a promising approach for tackling the robustness challenge under distribution shifts. However, the lack of consistent settings and systematic studies in prior literature hinders thorough assessments of existing methods. To address this issue, we present TTAB, a test-time adaptation benchmark that encompasses ten state-of-the-art algorithms, a diverse array of distribution shifts, and two evaluation protocols. Through extensive experiments, our benchmark reveals three common pitfalls in prior efforts. First, selecting appropriate hyper-parameters, especially for model selection, is exceedingly difficult due to online batch dependency. Second, the effectiveness of TTA varies greatly depending on the quality and properties of the model being adapted. Third, even under optimal algorithmic conditions, none of the existing methods are capable of addressing all common types of distribution shifts. Our findings underscore the need for future research in the field to conduct rigorous evaluations on a broader set of models and shifts, and to re-examine the assumptions behind the empirical success of TTA. Our code is available at \url{https://github.com/lins-lab/ttab}.
Much of the causal discovery literature prioritises guaranteeing the identifiability of causal direction in statistical models. For structures within a Markov equivalence class, this requires strong assumptions which may not hold in real-world datasets, ultimately limiting the usability of these methods. Building on previous attempts, we show how to incorporate causal assumptions within the Bayesian framework. Identifying causal direction then becomes a Bayesian model selection problem. This enables us to construct models with realistic assumptions, and consequently allows for the differentiation between Markov equivalent causal structures. We analyse why Bayesian model selection works in situations where methods based on maximum likelihood fail. To demonstrate our approach, we construct a Bayesian non-parametric model that can flexibly model the joint distribution. We then outperform previous methods on a wide range of benchmark datasets with varying data generating assumptions.
We consider model selection for sequential decision making in stochastic environments with bandit feedback, where a meta-learner has at its disposal a pool of base learners, and decides on the fly which action to take based on the policies recommended by each base learner. Model selection is performed by regret balancing but, unlike the recent literature on this subject, we do not assume any prior knowledge about the base learners like candidate regret guarantees; instead, we uncover these quantities in a data-driven manner. The meta-learner is therefore able to leverage the realized regret incurred by each base learner for the learning environment at hand (as opposed to the expected regret), and single out the best such regret. We design two model selection algorithms operating with this more ambitious notion of regret and, besides proving model selection guarantees via regret balancing, we experimentally demonstrate the compelling practical benefits of dealing with actual regrets instead of candidate regret bounds.
Automated model selection is an important application in science and engineering. In this work, we develop a learning approach for identifying structured dynamical systems from undersampled and noisy spatiotemporal data. The learning is performed by a sparse least-squares fitting over a large set of candidate functions via a nonconvex $\ell_1-\ell_2$ sparse optimization solved by the alternating direction method of multipliers. Using a Bernstein-like inequality with a coherence condition, we show that if the set of candidate functions forms a structured random sampling matrix of a bounded orthogonal system, the recovery is stable and the error is bounded. The learning approach is validated on synthetic data generated by the viscous Burgers' equation and two reaction-diffusion equations. The computational results demonstrate the theoretical guarantees of success and the efficiency with respect to the ambient dimension and the number of candidate functions.
We propose an automated computational algorithm for simultaneous model selection and parameter identification for the hyperelastic mechanical characterization of human brain tissue. Following the motive of the recently proposed computational framework EUCLID (Efficient Unsupervised Constitutive Law Identitication and Discovery) and in contrast to conventional parameter calibration methods, we construct an extensive set of candidate hyperelastic models, i.e., a model library including popular models known from the literature, and develop a computational strategy for automatically selecting a model from the library that conforms to the available experimental data while being represented as an interpretable symbolic mathematical expression. This computational strategy comprises sparse regression, i.e., a regression problem that is regularized by a sparsity promoting penalty term that filters out irrelevant models from the model library, and a clustering method for grouping together highly correlated and thus redundant features in the model library. The model selection procedure is driven by labelled data pairs stemming from mechanical tests under different deformation modes, i.e., uniaxial compression/tension and simple torsion, and can thus be interpreted as a supervised counterpart to the originally proposed EUCLID that is informed by full-field displacement data and global reaction forces. The proposed method is verified on synthetical data with artificial noise and validated on experimental data acquired through mechanical tests of human brain specimens, proving that the method is capable of discovering hyperelastic models that exhibit both high fitting accuracy to the data as well as concise and thus interpretable mathematical representations.
Deep learning models have become essential in software engineering, enabling intelligent features like image captioning and document generation. However, their popularity raises concerns about environmental impact and inefficient model selection. This paper introduces GreenRunnerGPT, a novel tool for efficiently selecting deep learning models based on specific use cases. It employs a large language model to suggest weights for quality indicators, optimizing resource utilization. The tool utilizes a multi-armed bandit framework to evaluate models against target datasets, considering tradeoffs. We demonstrate that GreenRunnerGPT is able to identify a model suited to a target use case without wasteful computations that would occur under a brute-force approach to model selection.