Proposed a stochastic online learning algorithm with probabilistic graph feedback, matching lower bounds.
Shuai Li, Wei Chen, Zheng Wen et al.
Using compositional ReLU dictionaries, depth 2 squares the rate O(N^{-η}), while depth 3 reaches O(N^{-2α/d}) for Hölder functions.
Zuowei Shen, Haizhao Yang, Shijun Zhang
Graph-VRNN combines visual evidence and interaction dynamics, achieving basketball fifth-step error 0.024 and strong Soccer World likelihood.
Chen Sun, Per Karlsson, Jiajun Wu et al.
Learned Step Size Quantization achieves highest accuracy for 2-4 bit precision on ImageNet.
Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani et al.
Proposes Meta Reward Learning (MeRL) combining KL divergence strategies to improve exploration and generalization in sparse reward settings for semantic parsing.
Rishabh Agarwal, Chen Liang, Dale Schuurmans et al.
AutoQ employs hierarchical DRL to automatically assign per-kernel QBN, reducing inference latency by 54.06% and energy by 50.69%, with minimal accuracy loss.
Qian Lou, Feng Guo, Lantao Liu et al.
This paper proves shallow neural networks with parameters exceeding the square root of training data size achieve global convergence via gradient descent.
Samet Oymak, Mahdi Soltanolkotabi
Proposes adapter modules for parameter-efficient transfer in NLP, achieving near state-of-the-art performance on 26 tasks with only 3.6% of full fine-tuning parameters.
Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski et al.
Introduces generalized sliced-Wasserstein (GSW) distance using nonlinear projections, reducing computational cost and improving high-dimensional distribution matching.
Soheil Kolouri, Kimia Nadjahi, Umut Simsekli et al.
Proposes Conditioning by Adaptive Sampling (CbAS) for protein design, leveraging generative models to estimate conditional distributions efficiently.
David H. Brookes, Hahnbeom Park, Jennifer Listgarten
Proposes TRADES, a defense method balancing robustness and accuracy via classification-calibrated loss, outperforming SOTA in experiments.
Hongyang Zhang, Yaodong Yu, Jiantao Jiao et al.
Introduces a rate-distortion-perception function, deriving closed-form solutions for Bernoulli sources, revealing perceptual constraints raise compression costs.
Yochai Blau, Tomer Michaeli
Diverse mini-batch Active Learning using K-means clustering reduces labeled data needs.
Fedor Zhdanov
Deep neural networks achieve exponential approximation error decay for diverse functions, based on Kolmogorov-optimal bounds.
Dennis Elbrächter, Dmytro Perekrestenko, Philipp Grohs et al.
PyOD integrates over 20 classical and neural network-based outlier detection algorithms with a unified API, supporting scalable anomaly detection in large datasets.
Yue Zhao, Zain Nasrullah, Zheng Li
GNNs like GCN, GAT, GRN leverage message passing to learn complex graph structures, achieving state-of-the-art performance.
Jie Zhou, Ganqu Cui, Shengding Hu et al.
Proposes a deep neural dialogue recommendation system using HRED and REDIAL dataset, achieving 75% recommendation accuracy.
Raymond Li, Samira Kahou, Hannes Schulz et al.
Proposes a continuous rotation representation, improving neural network learning with error reduction by 6-14 times.
Yi Zhou, Connelly Barnes, Jingwan Lu et al.
PointPillars uses PointNet-based pillar encoding, achieving 62Hz speed with superior accuracy on KITTI benchmarks.
Alex H. Lang, Sourabh Vora, Holger Caesar et al.
Soft Actor-Critic (SAC) combines maximum entropy RL with automatic temperature tuning, achieving state-of-the-art sample efficiency and stability for robotics.
Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen et al.