NNs-playground

Deep Learning architectures for multiclass image classification problems + autoencoders for MNIST digit dataset.

In this repository you can find source code for various Deep Learning (NNs) architectures. The code is provided in the form of Python Notebooks ( .ipynb files) cover the following:

• Multiclass image classification using Intel Image Classification dataset. The available classifiers are the following:
  • Multilayer Perceptron (MLP), using TensorFlow and Keras
  • Support Vector Machines (SVMs), using scikit-learn wrapper for LIBSVM
• Image reconstruction tasks for MINST digit dataset using TensorFlow and Keras. The available autoencoder flavors are the following:
  • classic convolutional autoencoder for image reconstruction
  • denoising autoencoder in several variations, including an autoencoder trained to reconstruct the noisy image
  • autoencoder to reconstruct the next digit
  • autoencoder to reconstruct the sum (as two images), given two input digits (as images)

The code was developed alongside with the Course Neural Networks - Deep Learning the author attended during their 6th semester of studies at the Computer Science Department of AUTh. The interested reader can find a presentation that summarizes the work produced.

Results

Next-digit autoencoder

Adder autoencoder

For a comprehensive description of the adopted workflow, as well as the produced results, the interested reader can refer to the presentation file.