Invictus-AI/Ivictus_Self-Evolving-Neural-network

🧬 Ivictus — Self-Evolving Neural Network

A self-evolving neural network that redesigns its own architecture using evolutionary strategies, plus a live web dashboard to watch and control it from your browser.

Creator: Kashyap Khakhkhar

---

✨ Capabilities

🧠 Self-Evolving Architecture Search

• Genome-based architecture system — every network layout (layer count, units, activations, dropout, batch-norm, learning rate, optimizer, batch size) is encoded as a mutable genome.
• Evolutionary loop — each generation mutates and crosses over the population, evaluates every model on real training data, and keeps only the fittest (elitism).
• Gated feature routing — the genome evolves which 3 input features drive the first hidden layer, with the rest injected into deeper layers.
• Diversity pressure — duplicate and near-duplicate architectures are penalized so the population doesn't collapse onto one local optimum.
• Smart fitness — classification optimizes validation accuracy (not loss) to prevent memorization; complexity is gently penalized (Occam's razor).

🛡️ Built-In Safety Gates

• Hard, immutable ceilings on layers, units, and total parameters that evolution can never exceed.
• Failure gates reject NaN/Inf losses and exploded models.
• Wall-clock budget and stagnation detection halt runs gracefully.

🎛️ Live Web GUI (evo_gui.py)

• Start / stop evolutionary runs from the browser.
• Live fitness chart (best vs. average over generations).
• Champion genome inspector + population table with fitness bars.
• Upload your own dataset (CSV / TSV / TXT / NPY / NPZ).
• Start from a saved model, a custom architecture, or the current champion.
• Streams the engine console to the browser in real time.

📊 Training on Your Own Data (evolve_dataset.py)

• Loads CSV / TSV / TXT / NPY / NPZ datasets with no extra dependencies.
• Auto-detects regression vs. classification, z-scores features, and targets the column you choose.
• Synthetic data generator when no dataset is provided.

💾 Checkpointing & Resumability

• Full evolution state (population, history, best genome) saved to disk every generation.
• Automatically resumes from checkpoints; --reset starts fresh.
• Best model exported as .keras plus genome config as JSON.

---

🎯 Benefits

Benefit	Why it matters
No architecture hand-tuning	The network finds the best design for your data automatically.
Generalizes better	Validation-based fitness + diversity pressure reduce overfitting.
Safe to run unattended	Hard safety gates mean evolution can't blow up memory, time, or loss.
Beginner-friendly GUI	Watch evolution live, tweak parameters, upload data — no ML expertise required.
Fully offline-capable	Works with synthetic data or your own files; no external API needed.
Resume anytime	Long runs survive restarts thanks to per-generation checkpoints.

---

Usage:

• The project Kashyap-K/self-evolving-nn is one of the testing of this when this was in developing stage.

🚀 Quick Start

pip install -r requirements.txt

# CLI: quick evolution run (synthetic data)
python3 self_evolving_model.py

# More generations / bigger population
python3 self_evolving_model.py --generations 100 --pop-size 12

# Web dashboard
python3 self_evolving_model.py --gui          # http://localhost:5000

# Train on your own dataset
python3 evolve_dataset.py data.csv --target label --task auto

Optional: the text-understanding path (e.g. --text-dataset rotten_tomatoes) lazily uses the HuggingFace datasets library — install it only if you need it: pip install datasets

---

📂 Files

File	Purpose
self_evolving_model.py	The self-evolving neural network engine + CLI.
evo_gui.py	Live web dashboard for the evolution engine.
evolve_dataset.py	Dataset loader & training driver for your own data.
requirements.txt	Python dependencies.

---

📜 License

Created by Kashyap Khakhkhar. Free to use, modify, and share.