All work
ResearchTransfer Learning · Computer Vision · Jan 2025 – Jun 2025
Domain Adaptation
Making CNNs generalise across datasets with CORAL feature alignment.
0
Benchmarks
MNIST · MNIST-M · Caltech-256 · Office-31
CORAL
Alignment
correlation alignment
Transfer
Method
+ augmentation
A study of cross-domain image classification: CNN models trained and evaluated across MNIST, MNIST-M, Caltech-256, and Office-31, using CORAL loss and data augmentation for robust feature alignment and transfer learning to cut dataset bias and improve generalisation.
The problem
Models trained on one dataset degrade sharply on another with a different distribution — and re-labelling every new domain is expensive or impossible.
The solution
Engineered CNN classifiers with CORAL (correlation-alignment) loss plus augmentation to align source and target feature statistics, then applied transfer learning to reduce dataset bias and boost cross-domain generalisation.
System Architecture
Challenges solved
- 1Aligning feature distributions across very different domains.
- 2Reducing dataset bias without target-domain labels.
- 3Fair evaluation across four heterogeneous benchmarks.
Highlights
- CORAL alignment across four vision benchmarks.
- Transfer learning to cut dataset bias.
- Foundational to the cross-domain medical-imaging work.
Technology
PythonPyTorchscikit-learnCORALCNNs
Next project
Genome Assembly & SARS-CoV-2 Quantification
Bioinformatics · Genomics