Work on known operator learning, face anti-spoofing, attention architectures and low-resource NLP. Newest first.
arXiv preprint · under review
An approach for estimating the statistical risk of known operator learning, where selected network layers are fixed as known operators instead of being learned. Gives a quantitative way to predict test error from training-set size — and shows how embedding prior physics cuts the data a model needs.
ICPR 2022 — 26th International Conference on Pattern Recognition
A multi-head angular self-attention network that sharpens feature extraction for biometric spoof detection, reaching state-of-the-art performance on face anti-spoofing benchmarks.
Sensors (MDPI) · 21(8), 2799
A bi-directional feature pyramid network that combines spatial features with Fourier spectra for pixel-wise face anti-spoofing, improving robustness across print, replay and mask attacks by attending to both high- and low-level cues.
DICTA 2020 — Digital Image Computing: Techniques and Applications
An attentional angular-margin extension of DeepPixBiS for pixel-wise face anti-spoofing, tightening the decision boundary between genuine and presentation attacks.
FTC 2020 — Future Technologies Conference (Springer AISC)
Contextual embeddings paired with oversampling to handle severe class imbalance in Bangla part-of-speech tagging, raising accuracy for a low-resource language.
Happy to talk through the methods, share code, or discuss collaborations.