New method distinguishes predictive distribution estimators in high-dimensional inputs.
problem Difficulty in evaluating predictive distributions for high-dimensional inputs.
method Introduces dyadic sampling to focus on predictive distributions associated with pairs of inputs.
result Demonstrates efficient distinction of predictive distribution estimators in high-dimensional examples.
This paper examines how the choice of prior distribution affects likelihoods of out-of-distribution inputs in deep generative models.
problem Mismatch between prior and data distributions causes deep generative models to assign higher likelihoods to out-of-distribution inputs.
method Proposes using a mixture distribution as a prior to make likelihoods of out-of-distribution inputs more sensitive.
result A mixture prior lowers the out-of-distribution likelihood with respect to real image data sets.
XEnsemble improves DNN robustness against adversarial and out-of-distribution inputs.
problem Protecting DNN models from adversarial and out-of-distribution inputs.
method Diverse input denoising verifiers and disagreement-diversity ensemble learning.
result XEnsemble achieves high defense and detection success rates.
Bayesian neural networks simplified with input augmentation.
problem Uncertainty in deep learning models.
method Layer-wise input augmentation to induce uncertainty distributions.
result State-of-the-art performance in uncertainty representation.
Proposes training neural networks to predict uncertainty for out-of-distribution inputs.
problem Poor uncertainty predictions for out-of-distribution inputs limit model robustness.
method Generates pseudo-inputs in low-density regions and trains a Bayesian framework.
result Yields robust and interpretable uncertainty predictions.
Paper develops neural network for distribution regression.
problem Regression with probability measures.
method Develops a novel fully connected neural network (FNN) for distribution inputs.
result Almost optimal learning rates for distribution regression derived.
The paper explores how benign overfitting occurs in heavy-tailed input distributions.
problem Understanding overfitting in heavy-tailed input distributions.
method Analysis of maximum margin classifiers on unregularized logistic loss with gradient descent.
result Linear classifiers trained under certain conditions can asymptotically achieve the noise level as misclassification error.
The ACCRU framework improves probabilistic forecasts by capturing input-dependent uncertainty.
problem Uncertainty in deterministic predictions, especially for skewed and non-Gaussian errors.
method Neural network trained with a loss function balancing accuracy and reliability to learn input-dependent, non-Gaussian uncertainty distributions.
result Improves probabilistic forecasts relative to existing methods, capturing skewed and non-Gaussian errors.
In-N-Out improves model robustness to out-of-distribution data.
problem Learning robust models with few in-distribution labeled examples.
method Pre-training with auxiliary information and self-training with pseudolabels.
result In-N-Out outperforms auxiliary inputs or outputs alone on both in-distribution and OOD error.
Neural Processes (NPs) (Garnelo et al 2018a;b) approach regression by learning to map a context set of observed input-output pairs to a distribution over regression functions. Each function models the distribution of the output given an input, conditioned on the context. NPs have the benefit of fitting observed data ef…
FDN improves probabilistic regressors' adaptability to distribution shifts.
problem Overconfidence in modern probabilistic regressors under distribution shift.
method FDN uses input-conditioned distributions over network weights, trained with a Monte Carlo beta-ELBO objective.
result FDN produces predictive mixtures whose dispersion adapts to the input, providing shift-aware uncertainty.
Wasserstein gradient boosting predicts probability distributions for supervised learning.
problem Distribution-valued supervised learning where outputs are probability distributions.
method Fits a new weak learner to Wasserstein gradients of loss functionals of probability distributions.
result Superior performance in probabilistic prediction compared to existing methods.
This work bridges outlier and drift detection by comparing inputs to a part of the reference distribution.
problem Monitoring machine learning models to ensure they operate within their validated distribution.
method Comparing a set of inputs to a chosen part of the reference distribution.
result A new approach that bridges outlier detection and drift detection.
New approach detects out-of-distribution inputs without needing OOD samples.
problem Detecting incorrect classification of out-of-distribution inputs in deep neural networks.
method A one-class classifier trained on an early layer's output of the original classifier.
result Substantially better results compared to state-of-the-art approaches.
Bayesian autoencoder detects AI safety risk of out-of-distribution inputs.
problem Detecting unreliable predictions from AI models with different distributions.
method Probabilistic, unsupervised Bayesian variational autoencoder with posterior estimation.
result Effective detection of out-of-distribution inputs in both input and latent spaces.
Efficiently aggregating data from different sources is a challenging problem, particularly when samples from each source are distributed differently. These differences can be inherent to the inference task or present for other reasons: sensors in a sensor network may be placed far apart, affecting their individual meas…
Paper improves distributed mean estimation and variance reduction without relying on input norm.
problem Distributed mean estimation and variance reduction with large input norms.
method Quantization and lattice theory connection for improved error bounds.
result Output error bounds depend only on input distance, not norm.
Sharpe et al. proposed the idea of having an expected utility maximizer choose a probability distribution for future wealth as an input to her investment problem instead of a utility function. They developed a computer program, called The Distribution Builder, as one way to elicit such a distribution. In a single-perio…
Develops a framework to quantify uncertainties in multiple ML models.
problem Uncertainty in ML model predictions and model inputs.
method Develops a theoretical framework to decouple and transform uncertainties.
result Generates joint distribution of ML predictions considering uncertainties.
DOODLER detects out-of-distribution inputs by reconstructing in-distribution data.
problem Detecting real-world out-of-distribution inputs for deep learning models.
method DOODLER uses a Variational Auto-Encoder to reconstruct in-distribution data and identifies failures as out-of-distribution.
result DOODLER outperforms other OOD detection methods under similar constraints.
Deep learning models converge to Gaussian dynamics with mixed structured inputs.
problem Understanding neural network dynamics with complex input distributions.
method Extended hidden manifold model to Gaussian mixtures, analyzed via SGD.
result Learning dynamics with mixed inputs converge to Gaussian behavior.
New findings show learning deeper neural networks is hard even with Gaussian inputs and non-degenerate weights.
problem The computational complexity of learning neural networks, especially deeper ones.
method Smoothed analysis framework and local pseudorandom generators.
result Learning depth-3 ReLU networks under Gaussian input distribution is hard even if weight matrices are non-degenerate.
New approach models how explanations shift with distribution changes.
problem Model performance drops with changing input data distributions.
method Models explanation shifts and compares them to state-of-the-art techniques.
result Modeling explanation shifts better detects out-of-distribution behavior.
Neural networks are vulnerable to small adversarial perturbations. Existing literature largely focused on understanding and mitigating the vulnerability of learned models. In this paper, we demonstrate an intriguing phenomenon about the most popular robust training method in the literature, adversarial training: Advers…
Improved VAE model enhances uncertainty estimation for out-of-distribution samples.
problem VAEs assign higher likelihood to out-of-distribution inputs.
method INCPVAE integrates noise contrastive prior into VAEs for reliable uncertainty estimation.
result INCPVAE outperforms standard VAEs in uncertainty estimation for OOD inputs.
MINs learn inverse mappings for high-dimensional optimization problems.
problem Data-driven optimization with high-dimensional inputs and valid subsets.
method Model Inversion Networks (MINs) learn an inverse mapping from scores to inputs.
result MINs can scale to high-dimensional input spaces and handle both offline and active data.
Paper shows attackers can steal model weights with just noise inputs.
problem Model weight theft with minimal inputs.
method Used i.i.d. Bernoulli noise inputs to achieve high accuracy.
result Achieved high accuracy (96% for MNIST, 82% for KMNIST) with minimal inputs.
Generative models learn distributions, new method finds inputs matching desired conditional distributions.
problem Designing inputs that produce specific conditional distributions, not just points.
method Conditional Distribution Matching (CDM) and MLGD-F algorithm.
result MLGD-F reliably recovers inputs matching diverse user-specified conditional distributions.
Simplified identification methods for causal inference with arbitrary interventional distributions.
problem Estimating cause-effect relationships from data with experimental interventions.
method Using Single World Intervention Graphs and nested model factorization, we provide algorithms for identifying causal parameters from mixed observational and interventional distributions.
result Our algorithms are complete for certain types of interventional marginal distributions.
LLMs' explanations are often insufficient and vary with input distribution.
problem Evaluating the sufficiency of LLM explanations without predefined biases.
method Generalizing sufficiency to arbitrary explanations, using LLM's input beliefs, and introducing SCSuff metric.
result Explanation sufficiency can vary with input distribution and is weakly correlated with model size, accuracy, or output entropy.
We analyze the convergence of (stochastic) gradient descent algorithm for learning a convolutional filter with Rectified Linear Unit (ReLU) activation function. Our analysis does not rely on any specific form of the input distribution and our proofs only use the definition of ReLU, in contrast with previous works that …
We give a new algorithm for learning a two-layer neural network under a general class of input distributions. Assuming there is a ground-truth two-layer network y=Aσ(Wx)+ξ, where A,W are weight matrices, ξ represents noise, and the number of neurons in the hidden layer is no larger than the input or outp…
Improved linear regression for diverse data batches.
problem Learning from multiple heterogeneous data sources.
method Gradient-based algorithm for different input distributions.
result Significant reduction in the number of required batches and sample size.
Deep learning provides a powerful tool for machine perception when the observations resemble the training data. However, real-world robotic systems must react intelligently to their observations even in unexpected circumstances. This requires a system to reason about its own uncertainty given unfamiliar, out-of-distrib…
Exact bounds derived for neural network outputs with noisy inputs.
problem Bounding the output distribution of neural networks with random inputs.
method Applying ReLU NNs to derive bounds for general NNs, then using these to find exact error guarantees.
result Exact upper and lower bounds for the output distribution of neural networks with random inputs.
Recent work has shown that deep generative models can assign higher likelihood to out-of-distribution data sets than to their training data (Nalisnick et al., 2019; Choi et al., 2019). We posit that this phenomenon is caused by a mismatch between the model's typical set and its areas of high probability density. In-dis…
Paper finds how many neurons are needed to approximate histogram distributions.
problem How many neurons are needed to approximate a target probability distribution?
method Examined for uniform input distribution and histogram target distributions, using efficient neural net construction.
result Obtained a new upper bound on the number of required neurons, strictly better than previous bounds.
In linear regression we wish to estimate the optimum linear least squares predictor for a distribution over d-dimensional input points and real-valued responses, based on a small sample. Under standard random design analysis, where the sample is drawn i.i.d. from the input distribution, the least squares solution for…
Generative classifier derived from any discriminative classifier rejects illegal inputs.
problem Detecting and rejecting illegal inputs like adversarial examples and out-of-distribution samples.
method SDIM-logit: learns generative classifier from logits of any discriminative classifier, imposing statistical constraints.
result SDIM-logit inherits performance of base classifier without loss and can reject illegal inputs.
This paper improves test-time adaptation for distribution shifts using confidence maximization and input transformation.
problem Improving deep networks' performance on data shifted from the training distribution.
method Proposes a novel loss function combining confidence maximization and batch-wise entropy maximization with an input transformation module.
result Significantly improves robustness of pretrained networks to corruptions on benchmarks like ImageNet-C.
Δ-UQ uses anchoring to estimate uncertainty in models.
problem Estimating uncertainty in predictive models.
method Anchoring input into a tuple for uncertainty estimation.
result Δ-UQ outperforms baselines in various use-cases.
Locus scores predictions for risk, reducing large-loss events.
problem Deployment cost from inaccurate predictions, especially large losses.
method Distribution-free loss-scale reliability score using any predictive distribution.
result Reduces large-loss frequency compared to standard heuristics.
Paper introduces a novel framework for set input tasks in meta-learning.
problem Meta-learning problems with set inputs often require efficient summary networks.
method Prototype-oriented optimal transport (POT) framework to improve summary networks.
result Significantly improves summary statistics from sets in meta-learning.
CNNs learn about input data uncertainties, improving classification performance.
problem Errors in measurements are often neglected in ML, affecting classification performance.
method Constructed a model with known noise levels and used CNNs to learn about uncertainties.
result CNNs can incorporate information about input data uncertainties, improving classification performance.
The paper develops GP classifiers for noisy inputs in multi-class classification.
problem Input noise in real-world data affects supervised learning performance.
method Developed multi-class Gaussian Process classifiers using variational inference.
result The proposed methods outperform noise-ignoring GP classifiers in predictive distribution.
Likelihood-based generative models are a promising resource to detect out-of-distribution (OOD) inputs which could compromise the robustness or reliability of a machine learning system. However, likelihoods derived from such models have been shown to be problematic for detecting certain types of inputs that significant…
Paper proposes using generalized lambda distributions for stochastic simulators.
problem Uncertainty quantification with complex stochastic models is computationally challenging.
method Flexible generalized lambda distribution approximates response PDF, parameters are sparse polynomial chaos expansions.
result Local inference of response PDF at each point of experimental design using replicated model evaluations.
New methodology controls synthetic data bias for neural program synthesis.
problem Deep networks generalize poorly to certain data distributions when trained on synthetic examples.
method Proposes a new methodology to control and evaluate the bias of synthetic data distributions over programs and specifications.
result Training deep networks on controlled synthetic data distributions leads to improved cross-distribution generalization performance.