mklarqvist / machine-learning-from-scratch

Machine learning from scratch

Machine learning from scratch

By: M. D. R. Klarqvist (mdrk.io)

This repository contains a collection of Jupyter notebooks focusing on the understanding and intuition behind the fundamentals of machine learning. The notebooks are written in Python and are intended to be used as a learning resource for those who are new to the field of machine learning. The notebooks are designed to be self-contained and to provide a step-by-step guide to understanding the core concepts of machine learning. The text that goes with the notebooks are available at https://mdrk.io.

The notebooks are designed to be interactive and to provide a hands-on learning experience. The notebooks are intended to be accessible to those with a basic understanding of Python and mathematics.

Table of Contents

• Introduction to machine learning using linear regression

  • What is linear regression?
  • Defining a linear regression model
  • Solving for the best possible parameters using calculus
  • Optimizing our model fit: Manually calculating the partial derivatives of a loss function with respect to our parameters
  • Optimizing our model fit: Using gradient descent to train the model using these manually computed partial derivative equations
  • Using stochastic gradient descent and mini-batch stochastic gradient descent to train the model
  • Introduction to higher-order optimization methods exemplified by Newton's method

• Regularization and the bias-variance trade-off (part 1)

  • Understanding overfitting and underfitting, and the bias-variance trade-off
  • Introducing the train-test split
  • How to add a L2 penalty term to our linear regression model to constrain the model to generalize better to unseen data
  • Optimizing the ridge regression model by computing the gradient manually
  • Optional section: deriving the Normal Equation
  • Understanding the regularization constraint and the method of Lagrange multipliers
  • Changing the L2 penalty to an L1 penalty to perfom Lasso regression
  • Applying what we have learned so far to a real-world application of predicting house prices

• Introduction to non-linear models exemplified by logistic regression

  • What is logistic regression?
  • Defining a logistic regression model
  • Understanding log odds and odds ratios and how they relate to logistic regression
  • Understanding logistic regression as a linear model with a non-linear activation function
  • Solving for the best possible parameters using calculus
  • Optimizing our model fit: Manually calculating the partial derivatives of a loss function with respect to our parameters
  • Optimizing our model fit: Using gradient descent to train the model using these manually computed partial derivative equations
  • Using stochastic gradient descent and mini-batch stochastic gradient descent to train the model

• Automatic differentiation (part 1)

  • Introducing dual numbers, their relationship to the derivative, and why they are important
  • A brief overview of the two main automatic differentation approaches
  • Implementing an automatic differentation program in Python from scratch
  • Verifying that the results are correct
  • Highlighting the pros and cons of the two main autodifferentiation approaches

• Automatic differentiation (part 2)

  • First we will revisit forward-mode automatic differentiation and look at function compositions and the chain rule
  • Breaking down the sequence of elementary operations into lists
  • Visualizing these lists graphically
  • Implementing these changes in Python from scratch
  • Introducing reverse-mode autodiff and an example
  • Implementing reverse-mode autodiff in Python from scratch

• Optimizers in Deep Learning

  • Covers several optimizers with math, intuition, and implementations from scratch in Python
  • SGD with Momentum
  • SGD with Nesterov Accelerated Gradient
  • Adam
  • RMSprop
  • Adagrad
  • Adadelta
  • Adamax
  • Nadam
  • AdamW

• Interesting Functions for Testing Optimization Methods

  • A series of interesting 2D functions to test optimize

• Stepping into the world of neural networks

  • What is a neural network?
  • Defining a neural network model
  • Understanding the feedforward process
  • Understanding the backpropagation process
  • Implementing linear regression neural network from scratch
  • Using automatic differentiation to train the neural network using backpropagation
  • Extending the neural network to include non-linear activation functions
  • Training a logistic regression neural network from scratch

More topics TBD.