Replication and discussion of experiments from the papers "Importance Weighted Autoencoders" ( https://arxiv.org/abs/1509.00519 ) and "Tighter Variational Bounds are Not Necessarily Better" ( https://arxiv.org/abs/1802.04537 ) on importance weighted variational autoencoders (IWAEs), with a few additional original experiments. Full description of the project is available in the pdf IWAE_replication.pdf. This was a joint work with Jamie Lee and Nicholas Pezzotti in the context of the University of Cambridge's MPhil in Machine Learning and Machine Intelligence .
The code was run using Tensorflow version 2.4.1 and Numpy version 1.19.5
Our experiments are run on the MNIST, fixed binarization MNIST, Fashion MNIST and Omniglot datasets.
An example of an experiment (training an autoencoder with the IWAE loss function on the fixed binarization dataset) can be found in experiment_example.py . We ran our experiments on Google's Colab; a few commands are specific to this environment.