Skip to content

Machine Learning Basics: Artificial Neural Networks, Generative Modeling, Boltzmann Machines, GANs

License

Notifications You must be signed in to change notification settings

dkomni/machine-learning-overview

Repository files navigation

An Overview of Machine Learning Techniques

Basics of machine learning techniques are presented along with a restricted Boltzmann machine (RBM) for generative modeling on Fashion-MNIST using Python. Also, a generative adversarial network (GAN) is implemented that generates synthetic data that resemble MNIST handwritten digits. This work was conducted at the context of my thesis preparation titled 'Quantum Computing for Generative Modeling and Applications' in Electrical & Computer Engineer at Technical University of Crete (TUC), Greece.

Overview

The pdf file Machine-Learning-Overview-DKomninos included in this repository presents some general Artificial Intelligence (AI) and Machine Learning (ML) concepts with thorough mathematics. Several basics such as neural networks, gradient descent and the backpropagation algorithm are discussed, along with comments on hyperparameters, datasets and training. Also, a general idea about generative adversarial neural networks (GANs) is introduced, arriving at the investigation of restricted Boltzmann Machines (RBMs) for generative modeling. A RBM is trained on the Fashion-MNIST dataset using the Contrastive Divergence algorithm and learns to generate synthetic data (reconstruct samples) that resemble the ones of the Fashion-MNIST dataset. The RBM may be further improved, as this is an implementation that tackles with aspects that Hinton presents in his guide. Moreover, a GAN is trained to generate synthetic handwritten digits from totally random inputs.

Notes

Whether using Windows, macOS or Linux, it is recommended that you create a virtual environment for Python and install the required packages within it. In this way, you isolate your operating system from the new installations and prevent possible compatibility issues. Also, make sure that you install the Jupyter kernel, which can be used to run Jupyter commands inside the virtual environment.

Special thanks to my friend Antonios Kastellakis for providing insights on RBMs and helped me with my introduction on generative modeling. You may find his thesis 'Analog and Digital Quantum Neural Networks: Basic Concepts and Applications', where he trains a RBM on the MNIST dataset of hand-written digits at: http://dimitrisangelakis.org/theses/

Some insight on the notebooks

01_MachineLearning_Overview

Implementation of simple neural networks, backpropagation algorithm and some data fitting using pure Python.

02_MachineLearning_Keras

Implementation of neural networks and tests on hyperparameters using Keras package from Tensorflow.

03_RBMs

Implementation of a restricted Boltzmann machine (RBM) for generative modeling on the Fashion-MNIST dataset, trained with Contrastive Divergence algorithm and reconstructing data with Gibbs sampling, along with results and comments. Heuristics such as momentum and learning rate decrease at the end of training are employed following Hinton's practical guide for training RBMs: https://www.cs.toronto.edu/~hinton/absps/guideTR.pdf

04_GANs

A generative adversarial network (GAN) trained to generate synthetic handwritten digits from random noise input is presented. This is just a basic presentation. The model may be further optimized.

Python version: 3.10.6

Package Version
Tensorflow 2.11.0
Matplotlib 3.6.3
Numpy 1.24.2
ipykernel 6.21.1
silence-tensorflow 1.2.1