paper / mikeyk / Jun 9
The paper introduces Counterfactual Transport Flows, a framework that refines low-feedback trajectories in offline RL by retrieving higher-feedback trajectory neighbors from logged data and transporting the candidate toward them in latent space. A refinement strength parameter controls the trade-off between preserving original behavior and applying stronger improvement. The method achieves interpretable, instance-specific refinement on D4RL benchmarks without requiring extrapolation beyond the offline data distribution.
offline-reinforcement-learningtrajectory-refinementcounterfactual-inferencemachine-learningpolicy-improvement
“Counterfactual Transport Flows improve offline RL policy without extrapolating beyond the offline data distribution.”
paper / mikeyk / Jun 6
Nd₂Be₂GeO₇, a frustrated Shastry-Sutherland magnet, exhibits a thermodynamic phase transition at 250 mK with no detectable static magnetism in zero field — a hallmark of hidden order. Uniquely, this hidden order is not suppressed by an applied magnetic field but is instead revealed only after field removal, producing magnetic Bragg peaks with strongly suppressed moments. µSR measurements confirm coherent (not glassy) spin fluctuations in the gigahertz range, indicating dynamic quantum coherence. The system is interpreted as sitting near a boundary between a quantum spin liquid and long-range magnetic order, where small perturbations can steer the ground state in either direction.
condensed-matter-physicsquantum-magnetismfrustrated-magnetsquantum-spin-liquidstrongly-correlated-electronsneutron-diffractionhidden-order
“A sharp specific-heat anomaly at 250 mK in Nd₂Be₂GeO₇ occurs without any detectable static magnetic order in zero field, consistent with a hidden-order phase transition.”
paper / mikeyk / Jun 3
DBSCAN, a widely-used density-based clustering algorithm, lacks any native mechanism to explain why a point is assigned as an inlier or outlier, or how robust that assignment is to perturbations. ExDBSCAN addresses this interpretability gap with a post-hoc, density-aware method that generates counterfactual explanations — showing what minimal changes would alter a point's cluster assignment. It uses a density-connected weighted graph and a physics-inspired repulsion-attraction model to produce multiple counterfactuals that are simultaneously valid, diverse, and proximal. Benchmarked against four baselines across 30 tabular datasets, ExDBSCAN achieves perfect validity while outperforming all competitors.
clusteringexplainabilitycounterfactual-explanationsunsupervised-learningdbscaninterpretable-ml
“Existing explainability methods for supervised machine learning are not directly applicable to clustering, creating an interpretability gap for algorithms like DBSCAN.”