Tutorial notebooks

Currently Playing: Tutorial 9: Deep Autoencoders (Part 1)

In this tutorial, we will take a closer look at autoencoders (AE). Autoencoders are trained on encoding input data such as images into a smaller feature vector, and afterward, reconstruct it by a second neural network, called a decoder. The feature vector is called the "bottleneck" of the network, since we aim to compress the input data into a smaller amount of features. This property is useful in many applications, in particular in compressing data or comparing images on a metric beyond pixel-level comparisons. Besides learning about the autoencoder framework, we will also see the "deconvolution" (or transposed convolution) operator in action for scaling up feature maps in height and width. Such deconvolution networks are necessary wherever we start from a small feature vector and need to output an image of full size (e.g. in VAE, GANs, or super-resolution applications). This notebook is part of a lecture series on Deep Learning at the University of Amsterdam. The full list of tutorials can be found at https://uvadlc-notebooks.rtfd.io. Link to notebook: https://uvadlc-notebooks.readthedocs.io/en/latest/tutorial_notebooks/tutorial9/AE_CIFAR10.html 00:00 Intro 01:33 Autoencoders 03:24 The Encoder 04:28 The Decoder 06:24 Combining Encoder and Decoder 07:54 Problems with MSE 09:33 Training the model 10:17 Comparing latent dimensionalities