Proposes a parameter-free Coin Betting-based meta algorithm (CBCE) with superior strongly adaptive regret bounds for changing environments.

Kwang-Sung Jun, Francesco Orabona, Rebecca Willett et al.

MMD-ResNet利用残差网络实现基于最大均值差异的非线性批次效应校准，有效减弱多源数据偏差。

Uri Shaham, Kelly P. Stanton, Jun Zhao et al.

Proposes linear classifier probes to monitor intermediate features, revealing monotonic increase in linear separability across layers.

Guillaume Alain, Yoshua Bengio

Proposes a VAE framework integrating DGDN decoder and CNN encoder for joint image, label, and caption modeling, achieving high efficiency and accuracy.

Yunchen Pu, Zhe Gan, Ricardo Henao et al.

Deep GRU-based multi-task learning significantly improves text recommendation accuracy, especially in cold-start scenarios, outperforming traditional models by up to 34%.

Trapit Bansal, David Belanger, Andrew McCallum

This study reveals the existence of bad local maxima in Gaussian Mixture Models and its impact on EM algorithm convergence.

Chi Jin, Yuchen Zhang, Sivaraman Balakrishnan et al.

Proposes Stein Variational Gradient Descent (SVGD), a particle-based Bayesian inference method minimizing KL divergence via kernel Stein distance.

Qiang Liu, Dilin Wang

Proposes two variance reduction methods—weighted M-estimator and control variates—for Bootstrap-based model aggregation, enhancing efficiency in distributed learning.

Jun Han, Qiang Liu

Proposes a differentially private deep learning training algorithm using Moments Accountant, achieving strong privacy guarantees with minimal accuracy loss.

Martín Abadi, Andy Chu, Ian Goodfellow et al.

Proposes LIME, a model-agnostic local explanation method using interpretable linear models to interpret complex black-box classifiers across diverse data types.

Marco Tulio Ribeiro, Sameer Singh, Carlos Guestrin

Unsupervised feature extraction using Time-Contrastive Learning and Nonlinear ICA, achieving first identifiability for nonlinear ICA.

Aapo Hyvarinen, Hiroshi Morioka

Proposes spectral regularization for statistical inverse learning, achieving minimax optimal convergence rates.

Gilles Blanchard, Nicole Mücke

Proposes a Bayesian optimization method with exponential convergence, avoiding auxiliary optimization and δ-cover sampling.

Kenji Kawaguchi, Leslie Pack Kaelbling, Tomás Lozano-Pérez

Provides a comprehensive mathematical framework for convolution, pooling, and transposed convolution size calculations, validated with experiments on CIFAR-10 and ImageNet.

Vincent Dumoulin, Francesco Visin

Proposes MF-GP-UCB, a multi-fidelity Gaussian process bandit algorithm that leverages low-cost approximations to improve black-box optimization efficiency.

Kirthevasan Kandasamy, Gautam Dasarathy, Junier B. Oliva et al.

Introduces Molecular Graph Convolutions, leveraging atom and bond features, achieving superior drug activity prediction with AUC up to 0.85.

Steven Kearnes, Kevin McCloskey, Marc Berndl et al.

ADVI automates variational inference via automatic differentiation, supporting non-conjugate models for large-scale Bayesian analysis.

Alp Kucukelbir, Dustin Tran, Rajesh Ranganath et al.

Proposes HOAG, an approximate gradient-based hyperparameter optimization method, outperforming traditional approaches in logistic and kernel ridge regression.

Fabian Pedregosa

Introduces KFC, a structured probabilistic approximation of the Fisher matrix for convolution layers, significantly accelerating training.

Roger Grosse, James Martens

Proposes a Gaussian Process-based estimation approach for optimization, eliminating parameter tuning, connecting GP-UCB and GP-PI, improving robustness.

Zi Wang, Bolei Zhou, Stefanie Jegelka