I build machine learning models that keep working when data is scarce, noisy or shifted.
Much of my work brings topology into deep learning: persistent homology, cell and simplicial complexes and hypergraphs give a network structure it can rely on when pixel-level cues fail. I pair this with hyperspectral and multimodal sensing, where I started, and apply it to crops and food, medical imaging, industrial inspection and radar.
I am a Research Fellow at La Trobe University with adjunct appointments at CSIRO and the Australian National University. I co-lead the CSIRO ML Workshop Series, supervise PhD and honours students, and received my PhD from Griffith University for work on 3D computer vision from hyperspectral images.
TopoOT chains optimal transport across persistence diagrams from different thresholds and filtrations, then trains a light head online on the features that stay stable. Up to +24.1% mean F1 over the next best method on 2D benchmarks.
TopoTTA replaces threshold heuristics with persistent homology on anomaly maps, producing pseudo-labels that keep defects connected. About 15% higher F1 on average across five 2D and 3D benchmarks.
Learns cubical multifiltrations over a joint range–Doppler, range–angle and angle–Doppler representation and feeds persistent H0 and H1 structure into radar segmentation end to end. On CARRADA it matches the strongest published methods (63.85% RD and 44.70% RA mIoU) with the best car IoU in both views.
A survey that connects spectral imaging with 3D vision: how spectral 3D data are acquired and represented, the learning methods built on them, benchmarks, and applications from agriculture to heritage. Computer Science Review, impact factor 17.5.
HyperRadar learns hypergraphs over radar returns that belong to the same object and aligns range–angle, range–Doppler and angle–Doppler views with unbalanced optimal transport, so scattered echoes are segmented as one pedestrian, cyclist or car.
EFN gives each embryo region graders use (trophectoderm, inner cell mass, blastocoel) its own expert subnetwork and fuses them with attention-gated weighting. Dice scores of 0.92, 0.91 and 0.90 on 2,344 annotated images, with cross-dataset validation.
A CLIP plant gate, YOLOv8 family detection, DINOv2 retrieval, allergen knowledge and live weather feed a rule-based risk score that a vision-language model explains. 91.2% top-1 accuracy on 26,978 plant images; the gate cut false acceptance of non-plant images from 83.8% to 0%.
Learnable hypergraphs and cross-view optimal transport for multi-view radar semantic segmentation.
Persistent-homology pseudo-labels for test-time adaptation in anomaly segmentation.
Hypergraph contrastive fusion of vibration signals and spectrograms for motor fault diagnosis.
RGB and point cloud fusion with attention-guided restoration for industrial anomaly detection.
Plus 7 hyperspectral computer vision repositories in MATLAB from my PhD and related projects.
All code github.com/alizia369