Jingwen Zhou

Implementation details can drastically alter the performance of training-free image detectors, with hyperparameters flipping conclusions about their effectiveness.

Switching scoring backends can boost domain-shift AUC by nearly 14 points, overshadowing the impact of pooling methods in training-free anomalous sound detection.

Transductive inference can boost zero-shot audio classification accuracy by up to 9.2 points, revealing the power of leveraging unlabeled data structures.

Simulate electrochemical interfaces with unprecedented accuracy and scale: this framework predicts both atomic forces and electron density distributions, unlocking new possibilities for understanding potential-driven phenomena.