AI Theory

Deep learning models’ robustness can be efficiently evaluated using a novel method, margin consistency, which leverages the correlation between input and logit margins for faster, accurate vulnerabili…

This paper proposes novel measures to detect and quantify confounding biases from observational data using causal mechanism shifts, even with unobserved confounders.

Stochastic gradient descent’s derivatives, crucial for hyperparameter optimization, converge to the solution mapping derivative; rates depend on step size, exhibiting O(log(k)²/k) convergence with van…

Derandomizing multi-distribution learning is computationally hard, but a structural condition allows efficient black-box conversion of randomized predictors to deterministic ones.

Denoising Diffusion Path (DDPath) uses diffusion models to dramatically reduce noise in attribution methods for deep neural networks, leading to clearer explanations and improved quantitative results.

DeNetDM uses network depth modulation to automatically debiase image classifiers without bias annotations or data augmentation, improving accuracy by 5%.

Deep linear networks implicitly regularize towards flat minima, with sharpness (Hessian’s largest eigenvalue) of minimizers linearly increasing with depth but bounded by a constant times the lower bou…

Deep Homomorphism Networks (DHNs) boost graph neural network (GNN) expressiveness by efficiently detecting subgraph patterns using a novel graph homomorphism layer.

New Perturbation Gradient losses connect expected decisions with directional derivatives, enabling Lipschitz continuous surrogates for predict-then-optimize, asymptotically yielding best-in-class poli…

Debiasing synthetic data generated by deep generative models enhances statistical convergence rates, yielding reliable results for specific analyses.

Data-faithful feature attribution tackles misinterpretations from unobservable confounders by using instrumental variables to train confounder-free models, leading to more robust and accurate feature …

Sublinear coresets for Poisson regression are developed, offering 1±ε approximation guarantees, with complexity analyzed using a novel parameter and domain shifting.

This paper proposes a novel MMD-based method for data distribution valuation, enabling theoretically-principled comparison of data distributions from limited samples, outperforming existing methods in…

Boost AI model robustness against adversarial attacks by creatively mixing training sample’s frequency amplitude with distractor images, focusing model learning on phase patterns, thus enhancing accur…

CSPG: a novel framework boosting Approximate Nearest Neighbor Search speed by 1.5-2x, using sparse proximity graphs and efficient two-staged search.

Score estimation, crucial for diffusion models, is computationally hard even with polynomial sample complexity unless strong distributional assumptions are made.

New definitions of differential privacy enable machine learning algorithms to credit sources appropriately, balancing data utility and copyright compliance.

Credal Learning Theory uses convex sets of probabilities to model data distribution variability, providing theoretical risk bounds for machine learning models in dynamic environments.

A new Covariate Shift Corrected Pearson Chi-squared Conditional Randomization (csPCR) test accurately assesses conditional independence even when data distributions vary between source and target popu…

This paper proposes a novel method to achieve optimal counterfactual fairness in machine learning models while minimizing predictive performance degradation.