ICML 2025 analysis rigorously quantifying when non-conservative force models (which predict forces directly) fail in molecular dynamics, demonstrating simulation instabilities and proposing hybrid architectures that capture speed benefits without sacrificing physical correctness.
ICML 2025 methodological paper introducing SPHNet, which uses adaptive network sparsification to overcome the computational bottleneck of tensor products in SE(3)-equivariant networks, achieving up to 7x speedup and 75% memory reduction on DFT Hamiltonian prediction tasks.
ICML 2025 paper proposing energy conservation metrics as critical diagnostics for machine learning interatomic potentials and introducing eSEN, a novel architecture designed to bridge the gap between test-set accuracy and real simulation performance on materials property prediction.
A NeurIPS 2021 method paper introducing Noise Contrastive Priors to address the VAE ‘prior hole’ problem, where standard Gaussian priors assign high density to regions of latent space that don’t correspond to realistic data, using energy-based models trained with contrastive learning to match the aggregate posterior.
Weiler et al.’s NeurIPS 2018 paper introducing 3D Steerable CNNs that achieve SE(3) equivariance through group representation theory and spherical harmonic convolution kernels, eliminating the need for rotational data augmentation and improving data efficiency for scientific applications with rotational symmetry like molecular and protein structures.
A provocative 2024 Nature Machine Intelligence paper challenging the assumption that invalid SMILES are failures, showing empirically that the ability to generate invalid outputs actually improves chemical language model performance by enabling quality filtering and providing richer training signals.
A complete guide to implementing modern Variational Autoencoders in PyTorch. Includes a copy-pasteable implementation, explanation of KL annealing to fix posterior collapse, and a deep dive into stable standard deviation parameterizations.
A PyTorch implementation enforcing strict Lyapunov stability guarantees on recurrent neural network controllers through Integral Quadratic Constraints, bridging 1990s robust control theory with modern deep reinforcement learning by solving semidefinite programs inside the gradient descent loop to provide mathematical certificates of safety.
We demonstrate that universal adversarial triggers can control both the topic and stance of GPT-2’s generated text, revealing security vulnerabilities in deployed language models and proposing constructive applications for bias auditing.
Discover how NEAT and HyperNEAT changed neuroevolution by automatically designing neural network architectures and scaling them through geometric patterns.
QuAC introduces a conversational QA dataset that models student-teacher interactions, creating context-dependent questions that test systems’ ability to understand dialogue and resolve references.
CoQA extends question answering beyond isolated questions to conversations that require context and reference understanding.