A comprehensive repository spanning deep learning fundamentals, computer vision, NLP, generative AI, and multi-agent systems — built and explored by Quach Thien Phu as part of his ML engineering journey at CSULB.
| # | Project | Category | Key Concepts | Tools |
|---|---|---|---|---|
| 01 | Building Your Deep Neural Network | Deep Learning Fundamentals | Forward/Backprop, Weights Init | NumPy, Python |
| 02 | Initialization Deep Learning | Deep Learning Fundamentals | Zero, Random, He Init | NumPy |
| 03 | Optimization Algorithms | Deep Learning Fundamentals | SGD, Momentum, Adam, RMSProp | NumPy |
| 04 | Regularization | Deep Learning Fundamentals | L2, Dropout, Early Stopping | NumPy |
| 05 | Gradient Checking | Deep Learning Fundamentals | Numerical gradient verification | NumPy |
| 06 | Planar Data Classification | Neural Networks | Binary classification, decision boundaries | NumPy, Matplotlib |
| 07 | Rain Prediction in Australia | Classical ML | EDA, Feature Eng, Classification | Sklearn, Pandas |
| 08 | Convolutional Neural Network | Computer Vision | Conv layers, Pooling, Filters | TensorFlow/Keras |
| 09 | Image Segmentation (U-Net V2) | Computer Vision | Encoder-Decoder, Skip Connections | TensorFlow/Keras |
| 10 | Sign Language Digits | Computer Vision | Multi-class classification | TensorFlow |
| 11 | MobileNet — Alpaca Classification | Transfer Learning | Fine-tuning, MobileNetV2 | TensorFlow/Keras |
| 12 | Face Recognition | Computer Vision | FaceNet, Triplet Loss, Embeddings | TensorFlow, OpenCV |
| 13 | Transformer (from Scratch) | NLP / Deep Learning | Attention, Positional Encoding, BERT | PyTorch |
| 14 | LLM ChatBot | Generative AI | Prompt Engineering, RAG basics | LangChain, OpenAI |
| 15 | Search Engine | Generative AI / IR | Vector Search, Embeddings | LangChain, FAISS |
| 16 | Agentic AI Hackathon | Agentic AI | Multi-agent orchestration | Fetch.ai, uAgents |
| 17 | Crew AI | Agentic AI | Role-based agents, task delegation | CrewAI |
| 18 | FellowAI | Agentic AI | Agent pipelines | LangGraph |
| 19 | NVIDIA — Accelerated Computing | Systems / Optimization | CUDA, GPU programming | NVIDIA toolkit |
Implementing a deep L-layer neural network from scratch using only NumPy.
- Built forward propagation, backward propagation, and parameter updates without any ML framework
- Implemented L-layer architecture with ReLU hidden layers and sigmoid output
- Applied to cat image binary classification task
- Key insight: Understanding the math behind gradient flow in deep networks
Input → [LINEAR → RELU] × (L-1) → LINEAR → SIGMOID → Output
Exploring how weight initialization affects training stability and speed.
| Method | Problem | When to use |
|---|---|---|
| Zero Init | Symmetry breaking fails | Never |
| Random Init | Exploding/vanishing gradients | Small networks |
| He Init | Optimal for ReLU layers | Deep networks |
Implementing and comparing gradient descent variants from scratch.
- Mini-batch GD — balances speed vs. accuracy
- Momentum — dampens oscillations, accelerates convergence
- RMSProp — adaptive learning rates per parameter
- Adam — combines Momentum + RMSProp, the go-to optimizer
Techniques to reduce overfitting in deep neural networks.
- L2 regularization (weight decay) — penalizes large weights
- Dropout — randomly deactivates neurons during training
- Early stopping — monitors validation loss to stop training
- Result: Significant reduction in overfitting on test datasets
Verifying backpropagation is implemented correctly using numerical gradient approximation.
- Applies fraud detection as the domain problem
- Computes numerical gradients via finite differences and compares with analytical gradients
- Use case: Debugging custom loss functions and layers
Training a shallow neural network to learn non-linear decision boundaries.
- Visualized how hidden layer size affects decision boundary complexity
- Compared logistic regression vs neural network on non-linear data
- Key takeaway: Even 1 hidden layer can approximate complex functions
End-to-end ML pipeline to predict next-day rainfall using weather data.
- Extensive EDA with Matplotlib and Seaborn visualizations
- Feature engineering, handling missing values, encoding categoricals
- Trained and compared Logistic Regression, Decision Tree, Random Forest, XGBoost
- Evaluated with ROC-AUC, Precision-Recall curves
Implementing CNNs for image classification tasks.
- Built Conv → Pool → Flatten → Dense architectures in TensorFlow/Keras
- Explored filter visualizations to understand what each layer learns
- Applied data augmentation to improve generalization
Semantic segmentation using the U-Net architecture.
- Encoder path: progressively downsamples with conv + maxpool
- Decoder path: upsamples with transposed convolutions
- Skip connections: preserve spatial information lost during downsampling
- Applied to biomedical image segmentation
Input Image
↓
[Encoder: Conv → MaxPool] × 4 ← Captures "what"
↓
Bottleneck
↓
[Decoder: UpConv + Skip] × 4 ← Recovers "where"
↓
Segmentation Mask
Multi-class classification of sign language digit images (0–9).
- Built CNN in TensorFlow from raw image data
- Preprocessing pipeline: normalization, one-hot encoding
- Achieved strong accuracy on 10-class gesture recognition
Transfer learning with MobileNetV2 for custom image classification.
- Loaded pre-trained MobileNetV2 weights (trained on ImageNet)
- Froze base layers, added custom classification head
- Fine-tuned top layers on the alpaca dataset
- Key insight: Transfer learning achieves high accuracy with minimal data
Building a face recognition system using deep embeddings.
- Implemented FaceNet architecture producing 128-dim face embeddings
- Triplet Loss: trains model to minimize distance between same-identity pairs
- One-shot learning: recognizes faces from a single reference image
- Real-time inference with OpenCV face detection
Full Transformer architecture implementation in PyTorch.
# Core components implemented:
├── Multi-Head Self-Attention
├── Positional Encoding
├── Feed-Forward Networks
├── Layer Normalization
├── Encoder & Decoder stacks
└── Masked attention for autoregressive decoding- Built encoder-decoder Transformer following "Attention Is All You Need"
- Explored BERT-style pre-training concepts
- Foundation for understanding GPT, BERT, T5, and modern LLMs
Conversational AI chatbot powered by large language models.
- Integrated OpenAI API for response generation
- Built with LangChain for conversation memory and chain orchestration
- Prompt engineering for consistent persona and context retention
- Deployed as an interactive interface
Semantic search engine using vector embeddings.
- Generates dense embeddings for documents using sentence transformers
- Indexes with FAISS for fast approximate nearest-neighbor search
- Returns semantically relevant results (not just keyword matches)
- Stack: LangChain + FAISS + OpenAI Embeddings
Multi-agent system built at a hackathon using Fetch.ai's uAgents framework.
- Designed specialized agents with distinct roles and communication protocols
- Agents coordinate to complete complex multi-step tasks autonomously
- Explored decentralized agent architecture
Role-based multi-agent collaboration using CrewAI framework.
- Defined AI agents with specific roles (researcher, writer, analyst)
- Orchestrated task delegation between agents
- Built end-to-end pipelines with autonomous decision-making
Advanced agent pipelines using LangGraph for stateful workflows.
- Built directed acyclic graph (DAG) agent workflows
- Implemented conditional routing and human-in-the-loop checkpoints
- Explored state management across long-running agentic tasks
GPU-accelerated ML using NVIDIA CUDA and optimized libraries.
- Explored CUDA programming fundamentals for parallel computation
- Applied cuDNN-accelerated deep learning operations
- Performance benchmarking: GPU vs CPU training times
Phase 1 — Foundations Phase 2 — Vision & NLP Phase 3 — Modern AI
──────────────────── ─────────────────────── ──────────────────────
[x] Neural nets from scratch [x] CNNs & image segmentation [x] Transformers & LLMs
[x] Optimization algorithms [x] Transfer learning [x] RAG & vector search
[x] Regularization [x] Face recognition [x] Agentic frameworks
[x] Gradient checking [x] Sign language detection [x] Multi-agent systems
[x] Classical ML + EDA [x] Object classification [ ] CGEV thesis (ongoing)
# Clone the repository
git clone https://github.com/quachphu/Machine-Learning-Project.git
cd Machine-Learning-Project
# Install dependencies
pip install -r requirements.txt
# Launch any notebook
jupyter notebook <ProjectFolder>/<notebook>.ipynbCommon dependencies:
pip install numpy pandas matplotlib seaborn scikit-learn
pip install tensorflow keras torch torchvision
pip install langchain openai faiss-cpu opencv-python
pip install crewai fetch-uagents langgraph
