DeepLabv3 integrates multi-scale atrous convolution and global features, achieving 85.7% mIOU on Pascal VOC 2012 without DenseCRF.

Liang-Chieh Chen, George Papandreou, Florian Schroff et al.

End-to-end negotiation model with dialogue rollouts significantly improves agreement success and strategic behavior.

Mike Lewis, Denis Yarats, Yann N. Dauphin et al.

Analyzes the Hessian matrix of over-parameterized neural networks, revealing a two-part spectrum structure.

Levent Sagun, Utku Evci, V. Ugur Guney et al.

SmoothGrad adds noise to input images and averages gradients, sharpening sensitivity maps for clearer model explanations.

Daniel Smilkov, Nikhil Thorat, Been Kim et al.

Transformer uses solely attention mechanisms, achieving 28.4 BLEU on WMT 2014 English-German translation, with faster training and superior quality.

Ashish Vaswani, Noam Shazeer, Niki Parmar et al.

Deep RL from human preferences enables complex task learning with less than 1% interaction, using preference-based reward modeling.

Paul Christiano, Jan Leike, Tom B. Brown et al.

Proposes a multi-document extractive summarization framework combining sentence ranking, keyphrase extraction, and semantic-based ordering, outperforming SOTA on DUC 2004.

Mir Tafseer Nayeem, Yllias Chali

Large-batch synchronized SGD with linear scaling and warmup trains ResNet-50 on ImageNet in 1 hour on 256 GPUs, maintaining accuracy.

Priya Goyal, Piotr Dollár, Ross Girshick et al.

PointNet++ employs a hierarchical architecture with multi-scale neighborhood features, significantly improving point cloud classification and segmentation accuracy.

Charles R. Qi, Li Yi, Hao Su et al.

Proposes MADDPG, a multi-agent actor-critic algorithm with centralized critic and decentralized execution, addressing non-stationarity and high variance in cooperative and competitive tasks.

Ryan Lowe, Yi Wu, Aviv Tamar et al.

Submanifold Sparse Convolutional Networks maintain sparsity while achieving state-of-the-art performance with 50% less computation.

Benjamin Graham, Laurens van der Maaten

Layer-wise fine-tuning of pre-trained CNNs outperforms training from scratch in medical image analysis, with performance gains of 5-12%.

Nima Tajbakhsh, Jae Y. Shin, Suryakanth R. Gurudu et al.

Proposes Sinkhorn divergence with entropic smoothing and automatic differentiation for scalable, stable training of generative models, bridging Wasserstein and MMD.

Aude Genevay, Gabriel Peyré, Marco Cuturi

Using Gumbel-softmax and reinforcement learning, the study induces multi-symbol protocols with natural language-like compositionality and variability.

Serhii Havrylov, Ivan Titov

Proposes a local descriptor learning loss based on maximizing the margin between closest positive and negative samples, improving matching accuracy.

Anastasiya Mishchuk, Dmytro Mishkin, Filip Radenovic et al.

ORGAN combines GAN and reinforcement learning to bias sequence generation towards specific metrics, improving molecular and music sample quality.

Gabriel Lima Guimaraes, Benjamin Sanchez-Lengeling, Carlos Outeiral et al.

Proposes Cramér distance to fix biased Wasserstein gradients, improving generative model stability and diversity.

Marc G. Bellemare, Ivo Danihelka, Will Dabney et al.

Defines low impact AI via impact penalty functions and probabilistic world coarse-graining, balancing usefulness and safety.

Stuart Armstrong, Benjamin Levinstein

This paper proves SGD convergence for two-layer ReLU networks with identity mappings under Gaussian inputs, showing polynomial-time global optimality and improved performance.

Yuanzhi Li, Yang Yuan

Cross-Alignment enables style transfer in non-parallel text, achieving 78.4% accuracy in sentiment modification.

Tianxiao Shen, Tao Lei, Regina Barzilay et al.