Survey of transfer learning methods in deep reinforcement learning, analyzing algorithms, applications, and challenges.

Zhuangdi Zhu, Kaixiang Lin, Anil K. Jain et al.

TadGAN uses GANs for time series anomaly detection, achieving the highest average F1 score.

Alexander Geiger, Dongyu Liu, Sarah Alnegheimish et al.

Graph Neural Network-based object importance prediction accelerates large-scale planning, outperforming baseline partial grounding strategies.

Tom Silver, Rohan Chitnis, Aidan Curtis et al.

PPG introduces phased training with separate policy and value updates, boosting sample efficiency by ~30% on Procgen benchmarks.

Karl Cobbe, Jacob Hilton, Oleg Klimov et al.

Proposes DeepAL framework combining Bayesian sampling and batch strategies, reducing labeling costs by 30% while maintaining high accuracy.

Pengzhen Ren, Yun Xiao, Xiaojun Chang et al.

HiPPO framework optimizes memory updates via polynomial projections, achieving 98.3% accuracy on permMNIST.

Albert Gu, Tri Dao, Stefano Ermon et al.

This survey reviews reinforcement learning (RL) methods for combinatorial optimization, focusing on TSP, from classical algorithms to deep RL with attention mechanisms, highlighting performance improvements.

Yunhao Yang, Andrew Whinston

Proposes a precise bi-level optimization framework for adversarial injection attacks, significantly improving attack success and transferability, validated on real datasets.

Jiaxi Tang, Hongyi Wen, Ke Wang

BigBird introduces sparse attention with global, local, and random tokens, reducing complexity from quadratic to linear, enabling long sequence modeling.

Manzil Zaheer, Guru Guruganesh, Avinava Dubey et al.

PackIt environment evaluates geometric planning using evolutionary algorithm, challenging task dataset validates effectiveness.

Ankit Goyal, Jia Deng

Proposes Self-Predictive Representations (SPR) to improve pixel-based deep RL sample efficiency, surpassing human scores within 100k steps.

Max Schwarzer, Ankesh Anand, Rishab Goel et al.

SUNRISE introduces a unified ensemble framework with weighted Bellman backups and UCB exploration, boosting deep RL stability and efficiency.

Kimin Lee, Michael Laskin, Aravind Srinivas et al.

Concept bottleneck models enable high-level concept prediction and intervention, maintaining competitive accuracy.

Pang Wei Koh, Thao Nguyen, Yew Siang Tang et al.

Self-supervised policy adaptation (PAD) enables reinforcement learning agents to online adapt in unseen environments without reward signals, significantly improving generalization.

Nicklas Hansen, Rishabh Jangir, Yu Sun et al.

Proposes a random-initialized single-step PGD method for fast adversarial training, extending to Euclidean norm, reducing training time while maintaining robustness and interpretability.

Justin Goodwin, Olivia Brown, Victoria Helus

AGE framework excels in node clustering and link prediction using adaptive encoding and Laplacian smoothing.

Ganqu Cui, Jie Zhou, Cheng Yang et al.

Proposes a contrastive framework for multimodal VAEs, leveraging PMI differences to improve data efficiency and sample distinction, validated on MNIST-SVHN and CUB datasets.

Yuge Shi, Brooks Paige, Philip H. S. Torr et al.

Proposes linear attention reducing Transformer complexity from O(N²) to O(N), enabling fast long-sequence autoregressive inference.

Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas et al.

Proposes a nearest-neighbor based gradient estimator with sharp nonasymptotic bounds, applicable to dimension reduction, optimization, and disentanglement, outperforming existing methods.

Guillaume Ausset, Stephan Clémençon, François Portier

Proposes epistemic uncertainty-aware robust imitative planning (RIP) for detecting and recovering from distribution shifts in autonomous driving.

Angelos Filos, Panagiotis Tigas, Rowan McAllister et al.