Autoencoding Beyond Pixels Using a Learned Similarity Metric

Similar documents
GAN and Feature Representation. Hung-yi Lee

arxiv: v1 [cs.lg] 31 Dec 2015

GAN Frontiers/Related Methods

When Variational Auto-encoders meet Generative Adversarial Networks

Alternatives to Direct Supervision

Variational Autoencoders. Sargur N. Srihari

arxiv: v2 [cs.lg] 17 Dec 2018

Deep Generative Models Variational Autoencoders

GENERATIVE ADVERSARIAL NETWORKS (GAN) Presented by Omer Stein and Moran Rubin

Unsupervised Learning

Adversarially Learned Inference

Extracting and Composing Robust Features with Denoising Autoencoders

19: Inference and learning in Deep Learning

Implicit generative models: dual vs. primal approaches

(University Improving of Montreal) Generative Adversarial Networks with Denoising Feature Matching / 17

arxiv: v1 [cs.lg] 10 Nov 2016

Lecture 19: Generative Adversarial Networks

Bidirectional GAN. Adversarially Learned Inference (ICLR 2017) Adversarial Feature Learning (ICLR 2017)

Deep generative models of natural images

Auto-Encoding Variational Bayes

Autoencoders. Stephen Scott. Introduction. Basic Idea. Stacked AE. Denoising AE. Sparse AE. Contractive AE. Variational AE GAN.

Autoencoder. Representation learning (related to dictionary learning) Both the input and the output are x

Model Generalization and the Bias-Variance Trade-Off

Neural Networks for Machine Learning. Lecture 15a From Principal Components Analysis to Autoencoders

Generating Images with Perceptual Similarity Metrics based on Deep Networks

Lecture 21 : A Hybrid: Deep Learning and Graphical Models

Autoencoders, denoising autoencoders, and learning deep networks

arxiv: v1 [cs.lg] 28 Feb 2017

Auto-encoder with Adversarially Regularized Latent Variables

Denoising Adversarial Autoencoders

An Empirical Evaluation of Deep Architectures on Problems with Many Factors of Variation

Lab meeting (Paper review session) Stacked Generative Adversarial Networks

Neural Networks: promises of current research

Introduction to Generative Adversarial Networks

Auxiliary Deep Generative Models

JOINT MULTIMODAL LEARNING WITH DEEP GENERA-

arxiv: v2 [stat.ml] 21 Oct 2017

Toward better reconstruction of style images with GANs

arxiv: v1 [cs.cv] 4 Apr 2018

Stacked Denoising Autoencoders for Face Pose Normalization

Day 3 Lecture 1. Unsupervised Learning

Deep Generative Models and a Probabilistic Programming Library

Akarsh Pokkunuru EECS Department Contractive Auto-Encoders: Explicit Invariance During Feature Extraction

arxiv: v2 [cs.lg] 9 Jun 2017

Lab 8 CSC 5930/ Computer Vision

arxiv: v2 [cs.lg] 25 May 2016

Variational Autoencoders

DOMAIN-ADAPTIVE GENERATIVE ADVERSARIAL NETWORKS FOR SKETCH-TO-PHOTO INVERSION

DOMAIN-ADAPTIVE GENERATIVE ADVERSARIAL NETWORKS FOR SKETCH-TO-PHOTO INVERSION

Semi-Amortized Variational Autoencoders

Generative Models in Deep Learning. Sargur N. Srihari

arxiv: v1 [cs.lg] 22 Sep 2016

Tutorial on Keras CAP ADVANCED COMPUTER VISION SPRING 2018 KISHAN S ATHREY

arxiv: v1 [cs.cv] 24 Apr 2018

Generative Adversarial Networks (GANs)

CSC321: Neural Networks. Lecture 13: Learning without a teacher: Autoencoders and Principal Components Analysis. Geoffrey Hinton

ADVERSARIAL EXAMPLES FOR GENERATIVE MODELS

GAN Related Works. CVPR 2018 & Selective Works in ICML and NIPS. Zhifei Zhang

Unsupervised Learning of Spatiotemporally Coherent Metrics

arxiv: v1 [cs.cv] 21 Jan 2018

The Multi-Entity Variational Autoencoder

Deep Manga Colorization with Color Style Extraction by Conditional Adversarially Learned Inference

Generative Modeling with Convolutional Neural Networks. Denis Dus Data Scientist at InData Labs

X-GAN: Improving Generative Adversarial Networks with ConveX Combinations

Learning to generate with adversarial networks

Deep-Q: Traffic-driven QoS Inference using Deep Generative Network

Deep Tracking: Biologically Inspired Tracking with Deep Convolutional Networks

Deep Hybrid Discriminative-Generative Models for Semi-Supervised Learning

COMP 551 Applied Machine Learning Lecture 16: Deep Learning

Image Restoration Using DNN

Unsupervised Image-to-Image Translation Networks

Iterative Inference Models

arxiv: v3 [cs.ne] 6 Dec 2016

Introduction to GAN. Generative Adversarial Networks. Junheng(Jeff) Hao

Deep Feature Interpolation for Image Content Changes

Deep Learning for Visual Manipulation and Synthesis

IMPROVING SAMPLING FROM GENERATIVE AUTOENCODERS WITH MARKOV CHAINS

arxiv: v2 [cs.cv] 24 Dec 2018

Generative Adversarial Text to Image Synthesis

ECE 599/692 Deep Learning. Lecture 12 GAN - Introduction

arxiv: v1 [stat.ml] 3 Apr 2017

Deep Meditations: Controlled navigation of latent space

DEEP LEARNING OF COMPRESSED SENSING OPERATORS WITH STRUCTURAL SIMILARITY (SSIM) LOSS

Visual Recommender System with Adversarial Generator-Encoder Networks

arxiv: v1 [stat.ml] 7 Nov 2016

Flow-GAN: Bridging implicit and prescribed learning in generative models

Virtual Adversarial Ladder Networks for Semi-Supervised Learning

Deep Learning. Deep Learning. Practical Application Automatically Adding Sounds To Silent Movies

arxiv: v1 [cs.gr] 27 Dec 2018

arxiv: v2 [cs.cv] 6 Dec 2017

An Empirical Study of Generative Adversarial Networks for Computer Vision Tasks

ALICE: Towards Understanding Adversarial Learning for Joint Distribution Matching

Tutorial Deep Learning : Unsupervised Feature Learning

COMPARATIVE DEEP LEARNING FOR CONTENT- BASED MEDICAL IMAGE RETRIEVAL

LEARNING TO INFER ABSTRACT 1 INTRODUCTION. Under review as a conference paper at ICLR Anonymous authors Paper under double-blind review

Neural Networks for unsupervised learning From Principal Components Analysis to Autoencoders to semantic hashing

Energy Based Models, Restricted Boltzmann Machines and Deep Networks. Jesse Eickholt

Tackling Over-pruning in Variational Autoencoders

Gradient of the lower bound

Machine Learning. Deep Learning. Eric Xing (and Pengtao Xie) , Fall Lecture 8, October 6, Eric CMU,

Transcription:

Autoencoding Beyond Pixels Using a Learned Similarity Metric International Conference on Machine Learning, 2016 Anders Boesen Lindbo Larsen, Hugo Larochelle, Søren Kaae Sønderby, Ole Winther Technical University of Denmark, University of Copenhagen, Twitter 16 December 2016 Presented by: Kevin Liang

Introduction Deep models have scored impressive successes for discriminative models on large and diverse datasets. However, generative models still present some problems. Variational Autoencoder (VAE) generated faces

Current Practice: Pixel-wise Similarity Metrics Many unsupervised learning methods (eg Variational Autoencoders) evaluate performance by comparing reconstruction with the original input. Common Metric: Pixel-wise squared error Problem: Humans do not perceive images element-wise

Proposed: Feature-wise Similarity Metric Use higher-level and sufficiently invariant representations of images for comparison, and then learn a function for similarity measure Propose jointly training a Variational Autoencoder (VAE) and a Generative Adversarial Network (GAN) VAE Decoder and GAN Generator shared GAN discriminator is the feature-similarity metric

Variational Autoencoder (VAE) Two networks: Encoder and Decoder z Enc(x) = q(z x), x Dec(z) = p(x z) (1) Regularize with a prior over latent distribution p(z) Typically: z N (0, I) Loss to minimize: with L VAE = E q(z x) [ log p(x z)p(z) ] q(z x) = L pixel llike + L prior (2) L pixel llike = E q(z x)[log p(x z)] (3) L prior = D KL (q(z x) p(z)) (4)

Generative Adversarial Network (GAN) Two networks: Generator and Discriminator x Gen(z), y = Dis(x) [0, 1] (5) Loss to maximize/minimize wrt Dis/Gen: L GAN = log(dis(x)) + log(1 Dis(Gen(z))) (6) with x being a training sample and z p(z)

VAE/GAN In order to properly distinguish generated images from real ones, GAN discriminators must implicitly learn a rich similarity metric. Let Dis l (x) be the hidden representation of lth layer of the discriminator: where x Dec(z) p(dis l (x) z) = N (Dis l (x) Dis l ( x), I) (7) Replace the pixel-wise error with a feature-based one from the discriminator: L Dis l llike = E q(z x)[log p(dis l (x) z)] (8)

VAE/GAN Loss VAE/GAN Loss: L = L prior + L Dis l llike + L GAN (9) L Dis l llike : Content Error L GAN : Style error Decoder and Generator share parameters

Practical Considerations Limiting error signals to relevant networks Dis should not try to minimize L Dis l llike Do not backpropagate the error signal from L GAN to Enc. Weighting VAE vs GAN Balance reconstruction with fooling the discriminator when training Dec θ Dec + θdec (γl Dis l llike L GAN) (10) Discriminating based on samples from p(z) and q(z x) In addition to exposing the discriminator to the usual fake and real data, encode the real data and add reconstructions L GAN = log(dis(x)) + log(1 Dis(Dec(z))) + log(1 Dis(Dec(Enc(x)))) (11)

Summary/Algorithm

Models Being Compared 4 Models, all with the same architectures for the sake of parity VAE: Vanilla VAE (pixel-wise Gaussian observation model) VAE Disl : VAE with learned feature-wise similarity metric from a GAN VAE/GAN: Similar to VAE Disl, but Dec also optimized wrt L GAN GAN: Vanilla GAN

Introduction Samples VAEs and GANs VAE/GAN Experiments

Reconstructions Note: GAN is not included, as it is incapable of performing reconstruction

Introduction VAEs and GANs VAE/GAN Visual Attribute Vectors - VAE/GAN Experiments

Introduction VAEs and GANs Attribute Query - VAE/GAN VAE/GAN Experiments

Shortcomings As a method of training a VAE: Train VAE/GAN model, throw away the GAN, and evaluate the remaining VAE using pixel-wise log likelihood Far from competitive with plain VAE models (on the CIFAR-10 datatset) Semi-supervised: Unsupervised pre-training with the VAE/GAN followed by fine-tuning on a small number of labeled examples Does not reach state-of-the-art (Ladder Networks, Stacked What-Where Autoencoders)