Analyzes FedAvg's performance drop on non-IID data via Earth Mover's Distance; sharing small global data improves accuracy by 30%.

Yue Zhao, Meng Li, Liangzhen Lai et al.

Proposed Low-rank Multimodal Fusion method significantly reduces computational complexity in tasks like sentiment analysis.

Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan et al.

Proposes PRD, a distribution metric separating quality and coverage, improving evaluation of generative models.

Mehdi S. M. Sajjadi, Olivier Bachem, Mario Lucic et al.

Trained RNN models on CoLA dataset for grammatical acceptability, achieving 67.1% accuracy, below human performance (86.1%).

Alex Warstadt, Amanpreet Singh, Samuel R. Bowman

Proposes PETS, combining uncertainty-aware deep dynamics models with trajectory sampling, achieving high sample efficiency and asymptotic performance in model-based RL.

Kurtland Chua, Roberto Calandra, Rowan McAllister et al.

Proposes Anonymous Walk Embeddings (AWE), combining feature-based and data-driven methods, to enhance unsupervised graph classification with state-of-the-art accuracy.

Sergey Ivanov, Evgeny Burnaev

Multi-modal deep learning model achieves 85% accuracy in pedestrian intent prediction, enhancing autonomous driving safety.

Amir Rasouli, John K. Tsotsos

The study uses statistical physics to compute entropy and mutual information in deep neural networks, exploring the link between compression and generalization.

Marylou Gabrié, Andre Manoel, Clément Luneau et al.

AutoAugment uses reinforcement learning to automatically discover image augmentation policies, reducing error rates to 1.5% on CIFAR-10 and achieving 83.5% Top-1 accuracy on ImageNet.

Ekin D. Cubuk, Barret Zoph, Dandelion Mane et al.

AutoPruner integrates end-to-end training for filter pruning, achieving over 50% FLOPs reduction with less than 1.2% accuracy drop.

Jian-Hao Luo, Jianxin Wu

A learning-based framework optimizes tensor programs for deep learning, reducing manual tuning and achieving performance comparable to hand-tuned libraries.

Tianqi Chen, Lianmin Zheng, Eddie Yan et al.

Hierarchical neural story generation with fusion and multi-scale self-attention significantly improves coherence and relevance, outperforming baselines.

Angela Fan, Mike Lewis, Yann Dauphin

Detect image splicing via learned self-consistency using real photo datasets, achieving state-of-the-art performance.

Minyoung Huh, Andrew Liu, Andrew Owens et al.

Using Luong’s seq2seq model, automatically translating informal LaTeX math texts into formal Mizar language achieved 65.73% accuracy on test data.

Qingxiang Wang, Cezary Kaliszyk, Josef Urban

Online normalizer calculation reduces memory accesses for Softmax, boosting performance by 1.3x; Softmax+TopK fusion boosts it by 5x.

Maxim Milakov, Natalia Gimelshein

Proposed a hypothesis-only baseline for NLI, significantly outperforming majority class baselines.

Adam Poliak, Jason Naradowsky, Aparajita Haldar et al.

A greedy edge-pruning method reveals a SAT–UNSAT transition, with tunable targets scaling as N^0.6–0.8.

Jason W. Rocks, Henrik Ronellenfitsch, Andrea J. Liu et al.

Introduces Winogender schemas to evaluate gender bias in coreference systems, revealing systematic bias correlated with real-world occupational gender statistics.

Rachel Rudinger, Jason Naradowsky, Brian Leonard et al.

This paper advocates using WMT standard for BLEU reporting, introducing SACREBLEU to ensure parameter consistency and comparability.

Matt Post

Proposes NES-based black-box attack algorithms for query-limited, partial, and label-only models, achieving over 90% success on ImageNet and Google API with fewer queries.

Andrew Ilyas, Logan Engstrom, Anish Athalye et al.