These tasks were created as part of the Deep Learning course at the Ben-Gurion University of the Negev. Each project covers a different field: ANN, CNN, RNN, and GAN.
Deep learning algorithms have recorded significant achievements in diverse areas such as image recognition, Text analysis, and robotics. The course provides an overview of the areas and trends in the field of deep learning today. The course opens with an overview of the theoretical basis on which the field stands and then reviews the different types of networks that exist Today and their suitability for various tasks. The course also discussed various issues related to the use of these algorithms For "real world" problems and in-depth examples of applications in various fields.
The beginning of the course was based on Coursera's course of Andrew Ng. - "DeepLearning Specialization" (link) and covered all of its syllabus.
Assignments
The purpose of the project was to build a simple neural network “from scratch” and to obtain a deep understanding of the forward/backward propagation process (without using TensorFlow or Keras).
Assignment 3 | Report - RNN
Lyrics Generation using RNNs
The purpose of this task was to build a recurrent neural net and use it on a real-world dataset. We trained a neural network to generate lyrics based on a provided melody. In addition to practical knowledge of the “how to” of building the network, an additional goal was to face the challenge of integrating different sources of information into a single framework.
Facial Recognition using One-shot Learning
The purpose of the task is to successfully execute a one-shot learning task for previously unseen objects, based on the paper Siamese Neural Networks for One-shot Image Recognition. Given two facial images of previously unseen persons, the architecture has to successfully determine whether they are the same person. We experimented with building a convolutional neural net and using it on a real-world dataset and problem.
Tabular samples generation
This work has two sections:
-
Implementation of a simple GAN model for tabular data.
-
Implementaion of a modified GAN architecture that can infer the inner-working of a black-box model.