The paper explores information-theoretic assumptions in batch reinforcement learning, providing theoretical results and sample complexity lower bounds.

Jinglin Chen, Nan Jiang

Introduced EigenPooling method to enhance graph classification performance, validated on 6 benchmark datasets.

Yao Ma, Suhang Wang, Charu C. Aggarwal et al.

Introduces an exact GPU-accelerated algorithm for Convolutional NTK, achieving 77% accuracy on CIFAR-10, outperforming previous kernel methods by 10%.

Sanjeev Arora, Simon S. Du, Wei Hu et al.

Proposes a Bayesian Generative Active Deep Learning method, enhancing classification efficiency on datasets like MNIST.

Toan Tran, Thanh-Toan Do, Ian Reid et al.

Sparse Transformer reduces attention complexity to O(n√n), enabling modeling of sequences tens of thousands long with hundreds of layers.

Rewon Child, Scott Gray, Alec Radford et al.

SAGPool enhances graph classification using self-attention, achieving a 5% improvement on the D&D dataset.

Junhyun Lee, Inyeop Lee, Jaewoo Kang

Proposed Graph Wavelet Neural Network (GWNN) using wavelet transform for efficient graph convolution, achieving state-of-the-art accuracy on Cora and other datasets.

Bingbing Xu, Huawei Shen, Qi Cao et al.

This study introduces membership inference attacks on sequence-to-sequence models, showing limited leakage under standard conditions but potential risks in overfitted models.

Sorami Hisamoto, Matt Post, Kevin Duh

Proposes an unsupervised deep neural network framework combining L2 and L∞ norms for solving PDE forward/inverse problems, applied to 2D elliptic systems.

Leah Bar, Nir Sochen

Centripetal SGD (C-SGD) merges filters during training, enabling pruning without fine-tuning, reducing FLOPs by 60% with minimal accuracy loss.

Xiaohan Ding, Guiguang Ding, Yuchen Guo et al.

Integrating structured scene representations and object-centric policies with RL and MCTS significantly improves complex construction task performance.

Victor Bapst, Alvaro Sanchez-Gonzalez, Carl Doersch et al.

Proposed D-MPNN with edge-based message passing improves molecular property prediction, outperforming traditional descriptors and previous GNNs.

Kevin Yang, Kyle Swanson, Wengong Jin et al.

LAMB optimizer trains BERT in 76 minutes on TPUv3 Pod with batch size 32868, maintaining performance.

Yang You, Jing Li, Sashank Reddi et al.

Proves random features require exponential parameters to approximate a single ReLU neuron, highlighting fundamental limitations in high-dimensional spaces.

Gilad Yehudai, Ohad Shamir

Variational Adversarial Active Learning (VAAL) achieves new SOTA on CIFAR10/100 using VAE and adversarial networks.

Samarth Sinha, Sayna Ebrahimi, Trevor Darrell

Deep generative graph neural network (CVGAE) learns molecular energy functions, generating conformations closer to references than force fields, with higher diversity and speed.

Elman Mansimov, Omar Mahmood, Seokho Kang et al.

ImageNet-C/P show that higher clean accuracy does not guarantee robustness to common corruptions or perturbation-induced prediction flips.

Dan Hendrycks, Thomas Dietterich

Proposed FAILMAKER-ADVRL employs multi-agent adversarial RL to generate failure scenarios for rule-based agents, achieving over 99% failure rate in tests.

Akifumi Wachi

Proposes Truly PPO, integrating a new clipping function and trust region trigger to enhance stability and sample efficiency.

Yuhui Wang, Hao He, Chao Wen et al.

GCOMB combines GCN and Q-learning, enabling scalable, high-quality heuristics on billion-scale graphs, achieving 100x speedup with marginally better solutions.

Sahil Manchanda, Akash Mittal, Anuj Dhawan et al.