Adversarial training leads to large generalization gap, decomposed into bias and variance.
problem Understanding the large generalization gap in adversarially trained models.
method Bias-Variance decomposition of test risk as a function of adversarial perturbation radius.
result Bias increases monotonically with adversarial perturbation radius and is dominant in test risk.
Proposes counterfactual explainability for causal attribution, extending variance analysis methods.
problem Lack of mechanistic understanding in existing tools for explaining complex models.
method Extends global sensitivity analysis methods to causal explanations using directed acyclic graphs.
result Developed methods to estimate counterfactual explainability and applied to income inequality analysis.
Large batch sizes reduce gradient variance in DP-SGD, improving privacy.
problem Understanding why large batch sizes work in DP-SGD.
method Decomposed total gradient variance into subsampling and noise-induced variances, proving batch size independence in the limit.
result Large batch sizes reduce effective total gradient variance, improving privacy in DP-SGD.
Proposes a simple method to explain aleatoric uncertainty in neural networks.
problem Lack of transparent explanations for uncertainty estimates in AI models.
method Adapting a neural network with Gaussian output to estimate predictive variance and applying explainers to the variance output.
result The proposed method explains uncertainty more reliably than complex approaches and outperforms them in most settings.
Contrastive Divergence (CD) and Persistent Contrastive Divergence (PCD) are popular methods for training the weights of Restricted Boltzmann Machines. However, both methods use an approximate method for sampling from the model distribution. As a side effect, these approximations yield significantly different biases and…
Proposes RVP to address theoretical concerns of V-REx for OOD generalization.
problem Theoretical concerns about V-REx's motivation and utility.
method Risk Variance Penalization (RVP) modifies V-REx's regularization.
result RVP discovers a robust predictor and finds invariant predictors under certain conditions.
Study on geodesics on random hyperbolic surfaces, showing variance asymptotic to X log X.
problem Distribution of closed geodesics on random hyperbolic surfaces.
method Viewing surfaces as random points in moduli space, studying weighted counting function.
result Variance in large genus limit is asymptotic to X log X, with exceptions.
Neural networks exhibit unimodal variance with model complexity, improving generalization.
problem The classical bias-variance trade-off does not apply to neural networks, leading to better generalization with larger models.
method Measured bias and variance of neural networks, confirmed empirically and theoretically.
result Neural networks show unimodal variance, leading to a double descent risk curve.
We study the dynamics of correlation and variance in systems under the load of environmental factors. A universal effect in ensembles of similar systems under the load of similar factors is described: in crisis, typically, even before obvious symptoms of crisis appear, correlation increases, and, at the same time, vari…
Study compares imputation methods' effects on IML confidence intervals.
problem Missing data impacts IML interpretation and confidence intervals.
method Compared single vs multiple imputation methods on IML confidence intervals.
result Multiple imputation provides closer coverage to nominal than single imputation.
Optimizes portfolios with utility theory, diversification, and leverage.
problem Finding optimal portfolio allocation strategies.
method Utility theory, exponential and logarithmic utilities, compound probability distributions, maximum expected utility, generalized mean-variance.
result Enhanced portfolio allocation strategies with natural explanations.
This paper develops a new theory for ensemble learning beyond variance reduction.
problem Ensemble learning's effectiveness for stable estimators is not fully explained by variance reduction.
method Develops a general weighting theory for ensemble learning, formalizing ensembles as linear operators and introducing geometric and spectral constraints.
result Structured weights can outperform uniform averaging by reshaping approximation geometry and redistributing spectral complexity.
The paper explores how Shapley value for a feature can vary based on model outcomes and feature distribution.
problem The uniqueness of Shapley value in explaining model predictions.
method Analyzes the relationship between feature distribution and Shapley value, and compares Shapley values for different model outcomes.
result Shapley value for a feature depends on more than just its mean and can vary significantly based on model outcome.
This paper discusses a novel explanation for asymmetric volatility based on the anchoring behavioral pattern. Anchoring as a heuristic bias causes investors focusing on recent price changes and price levels, which two lead to a belief in continuing trend and mean-reversion respectively. The empirical results support ou…
This paper discusses an alternative explanation for the empirical findings contradicting the positive relationship between risk (variance) and reward (expected return). We show that these contradicting results might be due to the false definition of risk-perception, which we correct by introducing Expected Downside Ris…
Unified framework explains few-shot multimodal medical imaging performance.
problem Limited labeled data in rare diseases and low-resource settings.
method PAC learning, VC theory, PAC Bayesian analysis, information gain, Chain of Thought reasoning.
result Unified theoretical framework for few-shot multimodal medical imaging.
A trade-off exists between reconstruction quality and the prior regularisation in the Evidence Lower Bound (ELBO) loss that Variational Autoencoder (VAE) models use for learning. There are few satisfactory approaches to deal with a balance between the prior and reconstruction objective, with most methods dealing with t…
Transformer-based models overfit financial time series data, leading to increased prediction variance.
problem Forecast collapse of transformer-based models under squared loss in financial time series.
method Theoretical analysis and numerical experiments on high-frequency EUR/USD exchange rate data.
result Increased model expressivity in Transformer-based models leads to spurious fluctuations without reducing bias, resulting in higher prediction variance.
The paper analyzes continuous optimization for DAG structure learning and its limitations.
problem The performance of continuous structure learning approaches is not consistent after data standardization.
method Analysis of continuous optimization for DAG structure learning, focusing on equal and non-equal noise variances.
result Continuous structure learning approaches may not perform well after data standardization, especially with non-equal noise variances.
Prediction and explanation are key objects in supervised machine learning, where predictive models are known as black boxes and explanatory models are known as glass boxes. Explanation provides the necessary and sufficient information to interpret the model output in terms of the model input. It includes assessments of…
Improved KernelSHAP via linear regression for ML model interpretation.
problem Efficiently estimating Shapley values in model-agnostic settings.
method Revisiting KernelSHAP via linear regression, developing techniques for convergence and uncertainty.
result Original KernelSHAP incurs negligible bias for significant variance reduction.
This work explains the structural origins of attention sinks in LLMs.
problem Initial tokens disproportionately monopolize attention scores in LLMs.
method Traced to self-attention's value aggregation process and FFN layer activations.
result Attention sinks form due to variance discrepancy and dimension disparity.
Paper explains how tree ensembles improve predictions by smoothing and regulating smoothness.
problem Understanding why tree ensembles perform well despite their complexity.
method Interpreting tree ensembles as adaptive and self-regularizing smoothers.
result Ensemble trees make more smooth predictions than individual trees and adjust smoothness based on input dissimilarity.
GNCL algorithm controls diversity in deep ensembles.
problem Managing bias and variance in deep ensembles.
method Generalized bias-variance decomposition for arbitrary loss functions, leading to GNCL algorithm.
result Explicit control over ensemble diversity and smooth interpolation between independent and joint training.
Many methods to explain black-box models, whether local or global, are additive. In this paper, we study global additive explanations for non-additive models, focusing on four explanation methods: partial dependence, Shapley explanations adapted to a global setting, distilled additive explanations, and gradient-based e…
Survey of determinism issues in financial AI systems.
problem Vulnerabilities in reproducibility of financial AI systems.
method Literature review and first-party experiments on public financial datasets.
result Proposed a layered evaluation framework linking modality-specific metrics to audit readiness.
Analyzes error sources in global feature effect estimation methods.
problem Unexplored error sources in global feature effect estimation methods.
method Systematic, estimator-level analysis of bias and variance.
result Holdout data is theoretically cleanest, but estimation variance depends on sample size and model characteristics.
AXE evaluates explanations to avoid misleading Rashomon set model selection.
problem Evaluating explanations for Rashomon set models to avoid false selection.
method Proposed AXE method to evaluate explanation quality.
result AXE detects adversarial fairwashing with 100% success rate.
In the era of deep learning, understanding over-fitting phenomenon becomes increasingly important. It is observed that carefully designed deep neural networks achieve small testing error even when the training error is close to zero. One possible explanation is that for many modern machine learning algorithms, over-fit…
The paper proposes criteria and methods for evaluating and aggregating feature-based model explanations.
problem Lack of quantitative evaluation criteria for feature-based model explanations.
method Developed quantitative evaluation criteria (low sensitivity, high faithfulness, low complexity), devised a framework for aggregation, and derived a new aggregate Shapley value explanation function.
result A new aggregate Shapley value explanation function that minimizes sensitivity.
New theory shows how multi-head attention reduces variance and decorrelates outputs.
problem Understanding and optimizing multi-head attention in neural networks.
method Developed a statistical theory linking multi-head attention to ensemble Nadaraya-Watson estimators.
result MHA variance reduction depends on head decorrelation, not just head count.
New definition reveals encoding explanations that retain predictive power.
problem Challenges in evaluating and identifying encoding explanations.
method Developed a definition of encoding based on conditional dependence.
result Existing evaluation scores do not rank non-encoding explanations correctly, but STRIPE-X does.
GRANITE unifies feature-based explanation methods to reduce disagreement.
problem Disagreement among feature-based explanation methods.
method GRANITE partitions feature space into regions minimizing interaction and distribution influences.
result Unified and consistent feature explanations.
Differentially private algorithms protect model explanations from leaking training data.
problem Model explanations can leak training data, compromising privacy.
method Adaptive differentially private gradient descent algorithm to produce accurate, private explanations.
result Privacy amplification and reduction of overall privacy loss on explanation data.
We consider objective evaluation measures of saliency explanations for complex black-box machine learning models. We propose simple robust variants of two notions that have been considered in recent literature: (in)fidelity, and sensitivity. We analyze optimal explanations with respect to both these measures, and while…
New insights into how high-dimensional models handle covariate shifts.
problem Covariate shift in high-dimensional random feature regression.
method Exact high-dimensional asymptotics of random feature regression under covariate shift.
result Overparameterized models exhibit enhanced robustness to covariate shift.
Improves global counterfactual explanations for model recourse.
problem Inability to provide explanations beyond local instances.
method Investigates and improves Actionable Recourse Summaries (AReS) for global counterfactual explanations.
result Develops more efficient and interactive explainability tools.
New framework evaluates model explanations based on decision task improvement.
problem Evaluation of model explanations often misses practical value.
method Decision-theoretic framework quantifying three key values.
result Provides benchmarks and interprets human-AI decision support.
G-SHAP generates multiple types of explanations for machine learning models.
problem Understanding model predictions and their differences across groups.
method Generalization of SHAP method to produce additional types of explanations.
result G-SHAP produces explanations for classification, intergroup differences, and model failure.
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.
Study finds visual explanations do not significantly improve human accuracy or trust in model predictions.
problem Measuring the impact of visual explanations on human accuracy and trust in model predictions.
method Randomized controlled trial with image-based age prediction task, varying levels of explanation quality.
result Visual explanations do not significantly alter human accuracy or trust in the model.
Despite a growing literature on explaining neural networks, no consensus has been reached on how to explain a neural network decision or how to evaluate an explanation. Our contributions in this paper are twofold. First, we investigate schemes to combine explanation methods and reduce model uncertainty to obtain a sing…
Many proposed methods for explaining machine learning predictions are in fact challenging to understand for nontechnical consumers. This paper builds upon an alternative consumer-driven approach called TED that asks for explanations to be provided in training data, along with target labels. Using semi-synthetic data fr…
Managing large-scale transportation infrastructure projects is difficult due to frequent misinformation about the costs which results in large cost overruns that often threaten the overall project viability. This paper investigates the explanations for cost overruns that are given in the literature. Overall, four categ…
This research investigates reliable local explanations for machine listening models.
problem Generating reliable local explanations for machine listening models.
method Investigates the sensitivity of SoundLIME explanations to input perturbations and proposes a novel method for identifying suitable content types.
result SoundLIME explanations are sensitive to the content in occluded input regions, and the average magnitude of input mel-spectrogram bins is the most suitable content type for temporal explanations.
Defines explanations for classifier outcomes using causal concepts.
problem Understanding classifier outcomes in a causal context.
method Proposes a new definition of explanation based on causality, compares it with existing notions, and evaluates it experimentally.
result Experimental evaluation shows the new definition's effectiveness on financial datasets.
A cornerstone of human statistical learning is the ability to extract temporal regularities / patterns from random sequences. Here we present a method of computing pattern time statistics with generating functions for first-order Markov trials and independent Bernoulli trials. We show that the pattern time statistics c…
This research improves interpretability in sequential explanations using mental models.
problem Improving interpretability in sequential explanations between two parties.
method A reinforcement learning framework that selects explanations based on the explainee's mental model.
result Mental model-based policies increase interpretability over random selection in multiple sequential explanations.