The Evolution: From "DIAE" to "Interactive"
In my BSc graduation project, DIAE (Dynamic Input-Adaptive Ensemble), I successfully demonstrated that we can significantly improve performance by dynamically weighting models based on the input. We achieved 97.05% accuracy by letting a "captain" decide which model to trust.
However, even in DIAE, the models (ResNet, DenseNet, EfficientNet) were still working in isolation. They extracted features separately, and we only combined their wisdom at the very end.
"Why not extract the extracted data? Why can't they help each other while they are working?"
The Innovation: Breaking the Walls
I decided to innovate and take it further. I designed an Interactive Ensemble Architecture that enables real-time information exchange during feature extraction.
Instead of just a "captain" picking the best player, the players now pass the ball to each other during the game. If ResNet detects a strong edge, it can "tell" DenseNet, which might use that context to better understand a texture.
Architecture: The Two-Pass Strategy
Intra-Backbone Coupling
We use Cross-Stitch Units and Cross-Attention Bridges to mix features at multiple network depths. This allows feedback to be injected into each backbone before its next stage.
Competence-Aware Routing
Similar to DIAE, we use a router. But this router predicts coupling strengths per-stage. It decides how much collaboration is needed at each depth of the network.
Ongoing Research
This is an ongoing research project aimed at solving the "Sequential Bottleneck" of traditional ensembles. By enabling models to "see" what others discover, we are moving towards a truly collaborative AI system.
Early validation shows a promising accuracy on our test set, outperforming traditional late fusion by 2-3%. The goal is to refine this architecture to handle even more complex histopathology cases where subtle morphological patterns determine critical diagnoses.