🧠 NeuroStack — AI/ML Learning Journey

A structured collection of notebooks and scripts covering the core domains of modern AI/ML — from classical algorithms to cutting-edge deep learning architectures, built through hands-on implementations and progressive experiments.

---

📁 Project Structure

NeuroStack/
├── 📂 Artificial_Intelligence/          # Search, planning, logic, agents (coming soon)
├── 📂 assets/
│   └── 🖼️  NeuroStack.png               # Project banner image
├── 📂 Computer_Vision/                  # Image processing, CNNs, object detection (coming soon)
├── 📂 Deep_Learning/                    # ANN, RNN, LSTM, GRU, Encoder-Decoder (Jupyter Notebooks)
│   ├── 📓 P1.1.ipynb                    # ANN — Apple Classification using Step Function (Binary)
│   ├── 📓 P1.2.ipynb                    # ANN — Apple Classification using Sigmoid + Normalization
│   ├── 📓 P2.1.ipynb                    # RNN — Forward Pass without Training (Random Weights)
│   ├── 📓 P2.2.ipynb                    # RNN — With Training on 2 Sequences (MSE Loss)
│   ├── 📓 P2.3.ipynb                    # RNN — Sequence Prediction (Next Number)
│   ├── 📓 P2.4.ipynb                    # RNN — Sentiment Analysis on Movie Reviews
│   ├── 📓 P3.1.ipynb                    # LSTM — Sequence Prediction (Next Number)
│   ├── 📓 P3.2.ipynb                    # LSTM — Sentiment Analysis on Movie Reviews
│   ├── 📓 P4.1.ipynb                    # GRU — Sequence Prediction (Next Number)
│   ├── 📓 P4.2.ipynb                    # GRU — Sentiment Analysis on Movie Reviews
│   └── 📓 P5.1.ipynb                    # Encoder–Decoder — Sequence Reversal (PyTorch GRU)
├── 📂 Machine_Learning/                 # Supervised, unsupervised, regression (coming soon)
├── 📂 Natural_Language_Processing/      # Text processing, embeddings, language models (coming soon)
├── 📂 Transformer/                      # Transformer architecture from scratch (Python Scripts)
│   ├── 🐍 P1.py                         # Self-Attention — Q, K, V matrices + scaled dot-product attention
│   ├── 🐍 P2.py                         # Multi-Head Attention — split heads, per-head attention + concat
│   ├── 🐍 P3.py                         # Positional Encoding — sinusoidal PE added to word embeddings
│   ├── 🐍 P4.py                         # Transformer Decoder — masked self-attention + causal language model
│   ├── 🐍 P5.py                         # Text Summarization — BART (facebook/bart-large-cnn) via HuggingFace
│   ├── 🐍 P6.py                         # Sentiment Analysis — BERT multilingual fine-tuned model
│   ├── 🐍 P7.py                         # Image Classification — Vision Transformer (ViT) via HuggingFace
│   ├── 🐍 P8.py                         # Text Generation — GPT-2 with temperature sampling
│   ├── 🐍 P9.py                         # Layer Normalization — manual implementation from scratch (PyTorch)
│   └── 🐍 P10.py                        # Full Transformer — nn.Transformer with src/tgt forward pass
├── 📄 .gitignore
├── 📄 LICENSE
└── 📄 README.md

---

🚀 Topics Covered

🧩 Domain	📚 Topics
Artificial Intelligence	Search algorithms, CSP, game theory, agents
Machine Learning	Linear/Logistic Regression, SVM, Decision Trees, Clustering
Deep Learning	ANNs, RNNs, LSTMs, GRUs, Encoder-Decoder
Computer Vision	Image classification, segmentation, OpenCV
NLP	Tokenization, embeddings, sentiment analysis, seq2seq
Transformer	Self-Attention, Multi-Head Attention, Positional Encoding, BERT, GPT-2, ViT

---

📒 Deep Learning — Notebook Details

🔷 ANN — Artificial Neural Networks

📓 Notebook	🏷️ Topic	📝 Description
P1.1.ipynb	ANN	Binary classification of apples using Color & Weight with Step Function activation
P1.2.ipynb	ANN	Same as P1.1 but with Sigmoid activation and feature normalization

🔁 RNN — Recurrent Neural Networks

📓 Notebook	🏷️ Topic	📝 Description
P2.1.ipynb	RNN	Simple RNN forward pass on 2 sequences without any training (random weights)
P2.2.ipynb	RNN	Simple RNN trained on 2 sequences to predict a single numeric output (MSE loss)
P2.3.ipynb	RNN	RNN trained to predict the next number in a sequence (sequence learning)
P2.4.ipynb	RNN	Sentiment analysis on 3 movie reviews using Embedding + SimpleRNN (binary classification)

🧠 LSTM — Long Short-Term Memory

📓 Notebook	🏷️ Topic	📝 Description
P3.1.ipynb	LSTM	LSTM trained to predict the next number in a sequence
P3.2.ipynb	LSTM	Sentiment analysis on movie reviews using Embedding + LSTM

⚡ GRU — Gated Recurrent Unit

📓 Notebook	🏷️ Topic	📝 Description
P4.1.ipynb	GRU	GRU trained to predict the next number in a sequence
P4.2.ipynb	GRU	Sentiment analysis on movie reviews using Embedding + GRU

🔄 Encoder–Decoder

📓 Notebook	🏷️ Topic	📝 Description
P5.1.ipynb	Encoder–Decoder	Sequence reversal using GRU-based Encoder–Decoder architecture in PyTorch

---

⚙️ Transformer — Script Details

🔩 Built from Scratch (NumPy / PyTorch)

🐍 Script	🏷️ Topic	📝 Description
P1.py	Self-Attention	Computes Q, K, V projections and scaled dot-product attention using NumPy
P2.py	Multi-Head Attention	Splits Q/K/V into multiple heads, computes per-head attention, concatenates and projects output
P3.py	Positional Encoding	Sinusoidal positional encoding added to word embeddings using PyTorch
P4.py	Transformer Decoder	Causal language model using nn.TransformerDecoder with masked self-attention
P9.py	Layer Normalization	Manual layer norm from scratch — mean, variance, gamma & beta using PyTorch
P10.py	Full Transformer	End-to-end nn.Transformer with random src/tgt tensors and forward pass

🤗 Pretrained Models (HuggingFace Transformers)

🐍 Script	🏷️ Topic	📝 Description
P5.py	Text Summarization	Abstractive summarization using facebook/bart-large-cnn pipeline
P6.py	Sentiment Analysis	Star-rating sentiment classification using nlptown/bert-base-multilingual-uncased-sentiment
P7.py	Image Classification	Image classification using Vision Transformer google/vit-base-patch16-224
P8.py	Text Generation	Open-ended text generation using GPT-2 with temperature sampling

---

🚀 Getting Started

# Clone the repo
git clone https://github.com/AbhishekGiri04/NeuroStack.git

cd NeuroStack

# Install dependencies
pip install -r requirements.txt

# Launch Jupyter for Deep Learning notebooks
jupyter notebook

# Run Transformer scripts directly
python Transformer/P1.py

Open any .ipynb file in Jupyter and run cells top to bottom. Python scripts can be run directly from terminal.

---

🎓 Learning Outcomes

After exploring this project, you will understand:

✅ ANN Fundamentals — Forward pass, activation functions (Step, Sigmoid), weight updates
✅ RNN Architecture — Sequential data processing, hidden states, backprop through time
✅ LSTM Internals — Forget, input & output gates, long-range dependency handling
✅ GRU Mechanics — Simplified gating, reset & update gates, efficiency vs LSTM
✅ Encoder–Decoder — Sequence-to-sequence learning, context vectors, GRU-based decoding
✅ Self-Attention — Query, Key, Value projections and scaled dot-product attention
✅ Multi-Head Attention — Parallel attention heads, head splitting and output projection
✅ Positional Encoding — Sinusoidal encoding to inject position into embeddings
✅ Transformer Decoder — Causal masking, autoregressive decoding
✅ Layer Normalization — Manual implementation of mean, variance, gamma & beta
✅ Pretrained Models — BART, BERT, GPT-2, ViT via HuggingFace Transformers
✅ Loss Functions — MSE for regression, Binary Cross-Entropy for classification
✅ PyTorch & Keras — Model building, training loops, and evaluation in both frameworks

---

🛠️ Tech Stack

🖥️ Technology	⚙️ Purpose	📊 What's Covered
	Core Language	Data processing, model logic, training loops
	Deep Learning	Transformer, GRU Encoder-Decoder, Layer Norm from scratch
	Deep Learning	Keras-based RNN, LSTM, GRU, Embedding layers
	Pretrained Models	BART, BERT, GPT-2, ViT pipelines
	Notebooks	Interactive experimentation & visualization
	Numerics	Array ops, manual attention computation

---

🌟 Key Features

• 📓 11 Notebooks — Progressive implementations: ANN → RNN → LSTM → GRU → Encoder-Decoder
• 🐍 10 Python Scripts — Transformer internals from scratch + HuggingFace pretrained models
• 📚 Educational — Clean, well-commented code with step-by-step logic
• 🔬 Hands-On — Real training loops, loss tracking, and predictions
• 🔧 Modular — Each file is self-contained and independently runnable
• ⚡ Dual Framework — Covers both PyTorch and TensorFlow/Keras
• 🤗 Pretrained Models — BART, BERT, GPT-2, ViT via HuggingFace
• 💡 Learning-Focused — Progressive difficulty from binary classification to full Transformer

---

📦 Requirements

• Python 3.8+
• Jupyter Notebook
• PyTorch
• TensorFlow / Keras
• NumPy
• HuggingFace Transformers
• Pillow
• Requests

---

📞 Contact & Support

👤 Abhishek Giri — Creator & Maintainer

---

📄 License

This project is open source and available under the MIT License — see the LICENSE file for details.

---

🧠 Built with ❤️ for Learning AI & Deep Learning

Mastering Neural Networks from the ground up

© 2026 Abhishek Giri | NeuroStack