🏔ī¸ ALPINE: Adaptive Patch Locator Demonstration

📄 arXiv:2609.22323 🤗 Hugging Face Model Checkpoints đŸ’ģ GitHub Codebase & Reproducibility

Accompanying the paper "ALPINE: Adaptive Localization for Parameter- and Sample-Efficient Few-Shot Learning" by Neeraj Yadav (Independent Researcher).

ℹī¸ Note on Implementation & Scope: This interactive demo runs the ALPINE Gabor edge-energy and adaptive locator equations directly in your browser (JavaScript), using the exact parameter values from the trained canonical model checkpoint (temperature = 1.00723886, window_frac = 0.50, and the fixed 4-orientation Gabor filter bank). It shows how the locator decides where to place patch sampling centers on any image in real-time — it does not run the full few-shot classification pipeline. Full PyTorch inference, training code, and all 5-seed model weights are available on GitHub and the Hugging Face model repository.

Interactive Patch Localization on Any Image

Standard vision architectures (like Vision Transformers) sample rigid, unmoving rectangular grids, often wasting patch capacity on uninformative background. Upload any photo below (or click a quick-start example) to see how ALPINE dynamically gravitates patch sampling centers toward salient biological edge contours.

Click to upload any image or drag and drop here

PNG, JPG, WEBP formats supported

Quick-Start Examples: Albatross Cardinal Indigo Bunting

1. Fixed-Grid Baseline (Static)

Rigid default coordinates

2. ALPINE Adaptive Locator

Dynamically centered on features

3. Gabor Edge Energy Map E(y,x)

Biological 4-orientation sum

Real-Time Patch Displacement Coordinates:

Patch Baseline (cx, cy) ALPINE Adaptive (cx, cy) Displacement (Δ)
ALPINE: Adaptive Localization for Parameter- and Sample-Efficient Few-Shot Learning â€ĸ Neeraj Yadav (Independent Researcher, India) â€ĸ ORCID: 0009-0000-7847-0588 â€ĸ MIT License