Construct minimally distorted adversarial examples via formal verification, enhancing adversarial training robustness by 4.2x.

Nicholas Carlini, Guy Katz, Clark Barrett et al.

HDLTex employs hierarchical deep neural networks to improve multi-level text classification, outperforming traditional methods with up to 97.97% accuracy.

Kamran Kowsari, Donald E. Brown, Mojtaba Heidarysafa et al.

This paper systematically analyzes reproducibility issues in deep RL, highlighting the impact of randomness, environment variability, and implementation details.

Peter Henderson, Riashat Islam, Philip Bachman et al.

Meta-learning-based one-shot visual imitation enables robots to learn new skills from a single demonstration with high success rates.

Chelsea Finn, Tianhe Yu, Tianhao Zhang et al.

Width-(n+4) ReLU networks are universal in L1, while width efficiency has a polynomial lower bound.

Zhou Lu, Hongming Pu, Feicheng Wang et al.

Proves that over-parameterized shallow neural networks with quadratic activation have no spurious local minima, enabling gradient descent to find global optima efficiently.

Mahdi Soltanolkotabi, Adel Javanmard, Jason D. Lee

Proposes a spectral-norm normalized margin bound for neural networks, showing its correlation with Lipschitz constants and generalization error.

Peter Bartlett, Dylan J. Foster, Matus Telgarsky

Proposes graph-based sampling strategies to accelerate neural network collaborative filtering, achieving up to 30x speedup.

Ting Chen, Yizhou Sun, Yue Shi 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.

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.

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

Serhii Havrylov, Ivan Titov

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

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

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

Proposes MMD GAN, integrating adversarial kernel learning to enhance GMMN, achieving superior image quality with smaller batch sizes.

Chun-Liang Li, Wei-Cheng Chang, Yu Cheng et al.

Using projected gradient descent, the paper proves linear convergence for learning ReLUs with near-optimal sample complexity in high dimensions.

Mahdi Soltanolkotabi

Proposes three stable deep neural network architectures inspired by ODE stability theory, ensuring robustness in extremely deep models.

Eldad Haber, Lars Ruthotto

Prototypical Networks utilize class means in an embedding space with Euclidean distance for few-shot classification, achieving state-of-the-art results.

Jake Snell, Kevin Swersky, Richard S. Zemel

Proposes Hypergradient Descent for dynamic learning rate adjustment, reducing manual tuning by leveraging automatic differentiation.

Atilim Gunes Baydin, Robert Cornish, David Martinez Rubio et al.

Integrated Gradients satisfies axioms of sensitivity and implementation invariance, providing a simple, model-agnostic attribution method for deep networks.

Mukund Sundararajan, Ankur Taly, Qiqi Yan