Machine Learning from Scratch – MLP and t-SNE

This repository contains two independent, from-scratch implementations of fundamental machine learning techniques:

1. Multi-Layer Perceptron (MLP) - a feedforward neural network for classification tasks.
2. t-Distributed Stochastic Neighbor Embedding (t-SNE) - a nonlinear dimensionality reduction technique for visualization of high-dimensional data.

Each implementation is contained in its own folder with a dedicated Jupyter Notebook (.ipynb) and a detailed README explaining the algorithm, step-by-step code, and comparisons with established libraries.

---

Repository Structure

.
├── MLP/
│   ├── MLP.ipynb
│   ├── README.md
├── T-SNE/
│   ├── T-SNE.ipynb
│   ├── README.md
└── README.md

---

1. MLP (Multi-Layer Perceptron)

Goal: Implement a fully-connected feedforward neural network using only NumPy, from scratch.

Highlights:

• Forward propagation and backpropagation implemented manually.
• Binary cross-entropy loss function and ReLU/Sigmoid activations.
• Custom weight initialization (Glorot and He methods).
• Training loop with mini-batch gradient descent.
• Benchmarking against TensorFlow's Sequential MLP.

Results: Comparable accuracy and loss trends with TensorFlow, validating the correctness of the implementation.

For detailed explanation, see MLP/README.md.

---

2. t-SNE

Goal: Implement t-SNE from scratch to project high-dimensional datasets into 2D while preserving local structure.

Highlights:

• Pairwise similarity computation in high-dimensional space.
• Perplexity-based adaptive variance selection using binary search.
• Low-dimensional embedding updates with gradient descent, momentum, and gains.
• Early exaggeration to improve cluster separation.
• Comparison with scikit-learn's t-SNE in terms of output and runtime.

Results: Produces visually similar cluster patterns to sklearn's t-SNE, though significantly slower due to pure NumPy implementation.

For detailed explanation, see T-SNE/README.md.

---

Requirements

• Python 3.8+
• Jupyter Notebook
• Libraries:

  numpy
  pandas
  matplotlib
  seaborn  # optional, for visualization
  scikit-learn  # only for dataset loading or comparison
  tensorflow  # optional, for MLP benchmarking
  

---

Running the Notebooks

The notebooks have integrated google colab links- you can view and run them.