Explore unsupervised learning and autoencoding techniques in this comprehensive 57-minute lecture. Delve into topics such as generative models, input and latent space interpolation, conditional generative networks, and style transfer. Learn about super resolution, inpainting, and caption-to-image generation using Dall-e. Understand key concepts like energy-based models, reconstruction energies, and loss functionals. Examine various autoencoder architectures, including denoising, nearest neighborhood, and sparse autoencoders. Gain insights into under and over-complete hidden layers, and conclude with final remarks on the subject.
Unsupervised Learning - Autoencoding the Targets
Overview
Syllabus
– 2021 edition disclaimer
– Unsupervised learning and generative models
– Input space interpolation
– Latent space interpolation
– Conditional generative networks
– Style transfer
– Super resolution
– Inpainting
– Caption to image Dall-e
– Definitions: x, y, z
– Recap: conditional latent variable EBM
– Recap: energy function
– Softmin training recap → autoencoder via amortised inference
– Reconstruction energies
– Loss functional
– Under and over complete hidden layer
– Denoising autoencoder
– Nearest neighbourhood denoising autoencoder
– Sparse autoencoder
– Final remarks
Taught by
Alfredo Canziani
