EDEN-MobileViTv3-Custom-ImageNet300

EDEN-MobileViTv3-Custom-ImageNet300 — SOTA Optimized

Primary KPI: EAG (Energy-to-Accuracy Gradient) = 1.9376e-10 ΔAcc/ΔJoules

Abstract

This model is part of Project EDEN (Energy-Driven Evolution of Networks), implementing the E2AM (Energy Efficient Advanced Model) Framework. The goal is to shift AI benchmarking from pure accuracy to Green SOTA — maximising predictive power per Joule consumed.

Applied Technique: Phase 2 – Progressive Unfreezing + AMP (E2AM SOTA)

Profiling Environment

🟢 Green Delta Table

Comparing this model against the reference baseline (ResNet-50 equivalent)

A positive EAG means this model learns more per Joule than the baseline. A negative EAG indicates a trade-off where higher accuracy required more energy investment.

E2AM Algorithm — Applied Phases

Phase 1 – Zero-Overhead Initialization: Dataset pre-loaded into pinned System RAM to eliminate disk I/O power spikes.

Phase 2 – Progressive Unfreezing: Backbone frozen for the first E_unfreeze epochs (only the classification head trains). At E_unfreeze, all layers are unfrozen and the learning rate is decayed. Gradient accumulation over N micro-batches simulates large batch sizes without proportional VRAM cost, slashing power-draw spikes.

AMP (Automated Mixed Precision): torch.cuda.amp.autocast() halves GPU memory bandwidth, reducing energy per backward pass.

Sparse Regularisation: L1 penalty λ·Σ|W| applied to trainable weights, driving dead neurons to zero and enabling future pruning.

Training Statistics

📊 Training Visualizations

Accuracy & Energy over Training

Green = accuracy (left axis) · Orange dashed = cumulative energy (right axis)

Project-Wide Overview

All EDEN models: energy vs accuracy

Cite This Research

@misc{eden2025,
  title     = {Project EDEN: Energy-Driven Evolution of Networks},
  author    = {EDEN Research Team},
  year      = {2025},
  note      = {Hugging Face: Shanmuk4622},
  url       = {https://huggingface.co/Shanmuk4622}
}