03 · Training & Performance

Inside the superposition framework.

Quoryn's representation layer compresses overlapping disease features into a sparse distributed code. These metrics track how well that code reuses features across pathologies — the core hypothesis of the framework.

Train / Val Loss
LOSS · 0.0042

Convergence under contrastive + reconstruction loss

Feature Reuse Rate
μ · 0.72

Shared latent basis across disease classes

Predictive Uncertainty
σ · 0.041

MC-dropout posterior over 200 samples

Architecture

Pipeline · Backbone → Superposition → Heads

  1. 01EfficientNet-B3 · pretrained backbone
  2. 02Sparse Variational Bottleneck · 4096-D
  3. 03Superposition Routing · top-k attention
  4. 04Multi-label heads · 38 disease classes
  5. 05MC-dropout · uncertainty estimation
  6. 06Grad-CAM · pixel-level attribution
Framework Note

Why superposition?

In a standard classifier, each disease occupies a disjoint region of feature space. Real-world plant samples violate that assumption — multiple pathogens, environmental stressors, and nutrient deficiencies coexist in the same tissue. Quoryn instead trains the penultimate layer as a sparse distributed code, where overlapping pathologies are represented by shared basis vectors rather than competing labels. This produces calibrated multi-label probabilities, interpretable latent activations, and graceful behaviour on out-of-distribution samples.