SlideShare una empresa de Scribd logo
1 de 37
Descargar para leer sin conexión
Cutting edge generative models
London TensorFlow meetup, March 2019
Pierre Harvey Richemond
Table of contents
1. Background
2. New in TensorFlow 2.0
3. Dual use
4. Perspectives
1
Background
Background : everything old is new again
• Advent of ’modern’ deep learning (resnets, batchnorm) : 2015
(arbitrary depth + large scale training)
• If stabilized, deeper is better.
• Invention of neural network training rule : 1970
(Linnainmaa’s masters thesis on backpropagation, applied to
neural networks)
2
Deep vs shallow networks
7x7 conv, 64, /2
pool, /2
3x3 conv, 64
3x3 conv, 64
3x3 conv, 64
3x3 conv, 64
3x3 conv, 64
3x3 conv, 64
3x3 conv, 128, /2
3x3 conv, 128
3x3 conv, 128
3x3 conv, 128
3x3 conv, 128
3x3 conv, 128
3x3 conv, 128
3x3 conv, 128
3x3 conv, 256, /2
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 512, /2
3x3 conv, 512
3x3 conv, 512
3x3 conv, 512
3x3 conv, 512
3x3 conv, 512
avg pool
fc 1000
image
3x3 conv, 512
3x3 conv, 64
3x3 conv, 64
pool, /2
3x3 conv, 128
3x3 conv, 128
pool, /2
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
pool, /2
3x3 conv, 512
3x3 conv, 512
3x3 conv, 512
pool, /2
3x3 conv, 512
3x3 conv, 512
3x3 conv, 512
3x3 conv, 512
pool, /2
fc 4096
fc 4096
fc 1000
image
output
size: 112
output
size: 224
output
size: 56
output
size: 28
output
size: 14
output
size: 7
output
size: 1
VGG-19 34-layer plain
7x7 conv, 64, /2
pool, /2
3x3 conv, 64
3x3 conv, 64
3x3 conv, 64
3x3 conv, 64
3x3 conv, 64
3x3 conv, 64
3x3 conv, 128, /2
3x3 conv, 128
3x3 conv, 128
3x3 conv, 128
3x3 conv, 128
3x3 conv, 128
3x3 conv, 128
3x3 conv, 128
3x3 conv, 256, /2
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 256
3x3 conv, 512, /2
3x3 conv, 512
3x3 conv, 512
3x3 conv, 512
3x3 conv, 512
3x3 conv, 512
avg pool
fc 1000
image
34-layer residual
3
Why does it work now ?
• In between : GPUs, fibre broadband, handheld megapixel
cameras, automatic differentiation software.
• All contributed to the explosion of applications we see today.
• Deep networks tend to work better and automatically find good
features for data classification & explanation on natural data:
images/video and sound.
• Other domains like text, structured data (chemistry...) can work
too with careful network design.
4
Deep Learning breakthroughs - 2014
• VGG architecture
• DeepDream
• First formulation of GANs
• First formulation of VAEs
• Black-box Variational Inference
• Neural Turing Machines
• Attention mechanisms
5
Deep Learning breakthroughs - 2015
• Residual Networks
• Batch Normalization
• ADAM optimizer
• ELU activation function
• Neural style transfer
• Graph convolutional networks
• Visual question answering (Karpathy)
• Normalizing Flows
• Deep Q-learning on Atari games
• Keras, TensorFlow
6
Deep Learning breakthroughs - 2016
• Wide Residual Networks
• DenseNets
• WaveNet
• PixelCNN
• Asynchronous Advantage Actor Critic (A3C)
• AlphaGo
• Neural Machine Translation at scale
• PyTorch
7
Deep Learning breakthroughs - 2017
• Wasserstein GAN
• Progressive Growing of GANs
• Capsule Networks
• Restarted SGD - cyclical learning rates
• AlphaGo Zero (MCTS as policy improvement ; no human
knowledge involved)
• Distributional Reinforcement Learning
• Equivalence of softmax Q-learning and entropic policy gradients
• TensorFlow distributions, Pyro
8
Deep Learning breakthroughs - 2018
• Spectral Normalization for GANs
• Fast initialization of convolutional neural networks
• Deep Video Portraits
• Differentiable neural architecture search
• OpenAI Five
• Video-to-video synthesis
• StyleGAN
9
A success story : Generative models
• Generative modelling aims at automatically explain features,
and generate new instances, of datapoints in a dataset
• Two main approaches that both use 2 different neural networks
instead of 1.
• Applications of the field to deep learning are 4 years old
• A success story of engineering
10
Generative models : GANs
Figure 1: Generative Adversarial Networks (GANs). A generator
(counterfeinter) and a discriminator (police) network play an adversarial
game, whose Nash equilibrium is perfect replication of the data distribution.
11
Generative models : VAEs
Figure 2: A typical variational autoencoder (VAE) architecture. An encoder
pushes input data through a (low) dimensionality bottleneck, that learns
relevant features in a latent code (optimized probabilistically). The decoder
network attempts to reconstruct inputs.
12
Illustrations - 2015
Figure 3: MNIST digits toy dataset. Interpolation.
13
Illustrations - 2016
Figure 4: Interpolations varying the strength of a smile vector computed by
doing latent space averaging of labeled pictures. Courtesy of Tom White.
14
Illustrations - 2017
Figure 5: GAN-generated 1024*1024 portraits from Karras et al., 2017.
15
Progressive Growing of GANs - Results, late 2018
Figure 6: High-resolution generations from Karras et al., 2018. 16
New in TensorFlow 2.0
TensorFlow 2.0 - General layout
Figure 7: Eager (dynamic graphing) and more Pythonic.
17
TensorFlow 2.0 - Layers declaration
Figure 8: Object-oriented layers.
18
TensorFlow 2.0 - Functional
Figure 9: The ’functional’ API, from Keras, is available.
19
TensorFlow 2.0 - VAE example
Figure 10: The functional API can shorten code significantly !
20
Dual use
From fake news to un-real people
www.thispersondoesnotexist.com
A blog post by Kyle McDonald on how to detect fake portraits:
21
Some photorealistic transformations - 2018
Deep Video Portraits (Portrait reenactment)
Vid2Vid (Photorealistic video-to-video translation)
See also TacoTron for realistic speech generation.
22
OpenAI GPT-2 - Principle
• A generative model trained at scale
• A relatively ’simple’ architecture and prediction concept, but the
scaling makes all the difference (8 million webpages training set)
• While it can be used in a multitask context ( question answering,
text summarization, translation...), it really shines in text
generation.
• For the first time, OpenAI has declined releasing the weights of
the full trained model, citing risks
• This has caused an unprecedented stir in the (self-regulated)
machine learning research community
23
OpenAI GPT-2 - Coherent paragraph generation - 1
When initalized with the following prompt text :
’In a shocking finding, scientist discovered a herd of unicorns living
in a remote, previously unexplored valley, in the Andes Mountains.
Even more surprising to the researchers was the fact that the
unicorns spoke perfect English.’
the model generated as a follow-up...
24
OpenAI GPT-2 - Coherent paragraph generation - 2
’ “The scientist named the population, after their distinctive horn,
Ovid’s Unicorn. These four-horned, silver-white unicorns were
previously unknown to science. Now, after almost two centuries, the
mystery of what sparked this odd phenomenon is finally solved.
Dr. Jorge Pérez, an evolutionary biologist from the University of La
Paz, and several companions, were exploring the Andes Mountains
when they found a small valley, with no other animals or humans.
Pérez noticed that the valley had what appeared to be a natural
fountain, surrounded by two peaks of rock and silver snow.
Pérez and the others then ventured further into the valley. ‘By the
time we reached the top of one peak, the water looked blue, with
some crystals on top,’ said Pérez.
Pérez and his friends were astonished to see the unicorn herd. These
creatures could be seen from the air without having to move too
much to see them – they were so close they could touch their horns.”’
25
OpenAI GPT-2 : Test 1 - Neural news generation
GPT2 seed ’DeepMind should have been a UK champion says
ex-Google CFO who took it to America’:
Google executives Mark Zuckerberg and Sergey Brin,formerly head of
the social media giant, have been accused of lying to the US
government about the extent of their ties to the Kremlin.
26
OpenAI GPT-2 : Test 2 - Scientific Generation !
’We compute a stochastic scaling function for data with a weighted
average of the mean as the value for which to apply an Likert
distribution over the data in the gradient descent. We then evaluate
the prediction accuracy and find that the weights are not completely
uniform in this way. Instead, the weight of this residual model is a
weighted average of the data in the gradient descent on the average,
which is also very accurate.’
27
Towards post-truth economics
• Shattering the realism barrier of the uncanny valley
• Legislation will be required (don’t hold your breath)
• Content industries could be notably disrupted
• Is this a serious use case for blockchain tech ?
Is the marginal cost of content creation going to zero ?
28
Perspectives
Are we at AGI yet ?
3 different views :
1. Killer robots are coming, let’s ban research !
2. AI is just matrix multiplication...
3. AI/Deep Learning is in fact letting an optimization algorithm
automatically write code for you
29
’Science sans conscience’ : TBC...
Figure 11: Human DNA data embedding and clustering (UMAP), categorized
by ethnicity. BiorXiv 2018, ’Revealing multi-scale population structure in
large cohorts’.
30
Questions?
30
Thank you for your attention
Twitter @KloudStrife
www.deeplearningmathematics.com

Más contenido relacionado

La actualidad más candente

Monitoring AI with AI
Monitoring AI with AIMonitoring AI with AI
Monitoring AI with AI
Stepan Pushkarev
 
ML at the Edge: Building Your Production Pipeline with Apache Spark and Tens...
 ML at the Edge: Building Your Production Pipeline with Apache Spark and Tens... ML at the Edge: Building Your Production Pipeline with Apache Spark and Tens...
ML at the Edge: Building Your Production Pipeline with Apache Spark and Tens...
Databricks
 
Managing the Complete Machine Learning Lifecycle with MLflow
Managing the Complete Machine Learning Lifecycle with MLflowManaging the Complete Machine Learning Lifecycle with MLflow
Managing the Complete Machine Learning Lifecycle with MLflow
Databricks
 

La actualidad más candente (20)

Graph Gurus 21: Integrating Real-Time Deep-Link Graph Analytics with Spark AI
Graph Gurus 21: Integrating Real-Time Deep-Link Graph Analytics with Spark AIGraph Gurus 21: Integrating Real-Time Deep-Link Graph Analytics with Spark AI
Graph Gurus 21: Integrating Real-Time Deep-Link Graph Analytics with Spark AI
 
Plume - A Code Property Graph Extraction and Analysis Library
Plume - A Code Property Graph Extraction and Analysis LibraryPlume - A Code Property Graph Extraction and Analysis Library
Plume - A Code Property Graph Extraction and Analysis Library
 
Monitoring AI with AI
Monitoring AI with AIMonitoring AI with AI
Monitoring AI with AI
 
Unifying State-of-the-Art AI and Big Data in Apache Spark with Reynold Xin
Unifying State-of-the-Art AI and Big Data in Apache Spark with Reynold XinUnifying State-of-the-Art AI and Big Data in Apache Spark with Reynold Xin
Unifying State-of-the-Art AI and Big Data in Apache Spark with Reynold Xin
 
Building Interpretable & Secure AI Systems using PyTorch
Building Interpretable & Secure AI Systems using PyTorchBuilding Interpretable & Secure AI Systems using PyTorch
Building Interpretable & Secure AI Systems using PyTorch
 
Graph Gurus Episode 1: Enterprise Graph
Graph Gurus Episode 1: Enterprise GraphGraph Gurus Episode 1: Enterprise Graph
Graph Gurus Episode 1: Enterprise Graph
 
ML at the Edge: Building Your Production Pipeline with Apache Spark and Tens...
 ML at the Edge: Building Your Production Pipeline with Apache Spark and Tens... ML at the Edge: Building Your Production Pipeline with Apache Spark and Tens...
ML at the Edge: Building Your Production Pipeline with Apache Spark and Tens...
 
Big Data Analytics With MATLAB
Big Data Analytics With MATLABBig Data Analytics With MATLAB
Big Data Analytics With MATLAB
 
DevOps and Machine Learning (Geekwire Cloud Tech Summit)
DevOps and Machine Learning (Geekwire Cloud Tech Summit)DevOps and Machine Learning (Geekwire Cloud Tech Summit)
DevOps and Machine Learning (Geekwire Cloud Tech Summit)
 
Managing the Complete Machine Learning Lifecycle with MLflow
Managing the Complete Machine Learning Lifecycle with MLflowManaging the Complete Machine Learning Lifecycle with MLflow
Managing the Complete Machine Learning Lifecycle with MLflow
 
Quoc Le at AI Frontiers : Automated Machine Learning
Quoc Le at AI Frontiers : Automated Machine LearningQuoc Le at AI Frontiers : Automated Machine Learning
Quoc Le at AI Frontiers : Automated Machine Learning
 
Graph Analytics for big data
Graph Analytics for big dataGraph Analytics for big data
Graph Analytics for big data
 
Scaling AI in production using PyTorch
Scaling AI in production using PyTorchScaling AI in production using PyTorch
Scaling AI in production using PyTorch
 
Graph Gurus Episode 12: Tiger Graph v2.3 Overview
Graph Gurus Episode 12: Tiger Graph v2.3 OverviewGraph Gurus Episode 12: Tiger Graph v2.3 Overview
Graph Gurus Episode 12: Tiger Graph v2.3 Overview
 
MLOps - Build pipelines with Tensor Flow Extended & Kubeflow
MLOps - Build pipelines with Tensor Flow Extended & KubeflowMLOps - Build pipelines with Tensor Flow Extended & Kubeflow
MLOps - Build pipelines with Tensor Flow Extended & Kubeflow
 
Satwik Mishra resume
Satwik Mishra resumeSatwik Mishra resume
Satwik Mishra resume
 
Training at AI Frontiers 2018 - LaiOffer Data Session: How Spark Speedup AI
Training at AI Frontiers 2018 - LaiOffer Data Session: How Spark Speedup AI Training at AI Frontiers 2018 - LaiOffer Data Session: How Spark Speedup AI
Training at AI Frontiers 2018 - LaiOffer Data Session: How Spark Speedup AI
 
Big (chemical) data? No Problem!
Big (chemical) data? No Problem!Big (chemical) data? No Problem!
Big (chemical) data? No Problem!
 
Automated Production Ready ML at Scale
Automated Production Ready ML at ScaleAutomated Production Ready ML at Scale
Automated Production Ready ML at Scale
 
END-TO-END MACHINE LEARNING STACK
END-TO-END MACHINE LEARNING STACKEND-TO-END MACHINE LEARNING STACK
END-TO-END MACHINE LEARNING STACK
 

Similar a TensorFlow London: Cutting edge generative models

Similar a TensorFlow London: Cutting edge generative models (20)

Deeplearning in finance
Deeplearning in financeDeeplearning in finance
Deeplearning in finance
 
Evolution of Deep Learning and new advancements
Evolution of Deep Learning and new advancementsEvolution of Deep Learning and new advancements
Evolution of Deep Learning and new advancements
 
Deep learning: Cutting through the Myths and Hype
Deep learning: Cutting through the Myths and HypeDeep learning: Cutting through the Myths and Hype
Deep learning: Cutting through the Myths and Hype
 
A historical introduction to deep learning: hardware, data, and tricks
A historical introduction to deep learning: hardware, data, and tricksA historical introduction to deep learning: hardware, data, and tricks
A historical introduction to deep learning: hardware, data, and tricks
 
Introduction to the Artificial Intelligence and Computer Vision revolution
Introduction to the Artificial Intelligence and Computer Vision revolutionIntroduction to the Artificial Intelligence and Computer Vision revolution
Introduction to the Artificial Intelligence and Computer Vision revolution
 
Unraveling Information about Deep Learning
Unraveling Information about Deep LearningUnraveling Information about Deep Learning
Unraveling Information about Deep Learning
 
AI is Impacting HPC Everywhere
AI is Impacting HPC EverywhereAI is Impacting HPC Everywhere
AI is Impacting HPC Everywhere
 
Deep learning and computer vision
Deep learning and computer visionDeep learning and computer vision
Deep learning and computer vision
 
Artificial Intelligence, Machine Learning and Deep Learning
Artificial Intelligence, Machine Learning and Deep LearningArtificial Intelligence, Machine Learning and Deep Learning
Artificial Intelligence, Machine Learning and Deep Learning
 
APPLICATION OF CONVOLUTIONAL NEURAL NETWORK IN LAWN MEASUREMENT
APPLICATION OF CONVOLUTIONAL NEURAL NETWORK IN LAWN MEASUREMENTAPPLICATION OF CONVOLUTIONAL NEURAL NETWORK IN LAWN MEASUREMENT
APPLICATION OF CONVOLUTIONAL NEURAL NETWORK IN LAWN MEASUREMENT
 
APPLICATION OF CONVOLUTIONAL NEURAL NETWORK IN LAWN MEASUREMENT
APPLICATION OF CONVOLUTIONAL NEURAL NETWORK IN LAWN MEASUREMENTAPPLICATION OF CONVOLUTIONAL NEURAL NETWORK IN LAWN MEASUREMENT
APPLICATION OF CONVOLUTIONAL NEURAL NETWORK IN LAWN MEASUREMENT
 
A Survey of Deep Learning Algorithms for Malware Detection
A Survey of Deep Learning Algorithms for Malware DetectionA Survey of Deep Learning Algorithms for Malware Detection
A Survey of Deep Learning Algorithms for Malware Detection
 
An Extensive Review on Generative Adversarial Networks GAN’s
An Extensive Review on Generative Adversarial Networks GAN’sAn Extensive Review on Generative Adversarial Networks GAN’s
An Extensive Review on Generative Adversarial Networks GAN’s
 
[PR12] Inception and Xception - Jaejun Yoo
[PR12] Inception and Xception - Jaejun Yoo[PR12] Inception and Xception - Jaejun Yoo
[PR12] Inception and Xception - Jaejun Yoo
 
ProjectReport
ProjectReportProjectReport
ProjectReport
 
Neural Networks and Deep Learning for Physicists
Neural Networks and Deep Learning for PhysicistsNeural Networks and Deep Learning for Physicists
Neural Networks and Deep Learning for Physicists
 
HML: Historical View and Trends of Deep Learning
HML: Historical View and Trends of Deep LearningHML: Historical View and Trends of Deep Learning
HML: Historical View and Trends of Deep Learning
 
Tutorial on Deep Learning
Tutorial on Deep LearningTutorial on Deep Learning
Tutorial on Deep Learning
 
Deep Learning Representations for All - Xavier Giro-i-Nieto - IRI Barcelona 2020
Deep Learning Representations for All - Xavier Giro-i-Nieto - IRI Barcelona 2020Deep Learning Representations for All - Xavier Giro-i-Nieto - IRI Barcelona 2020
Deep Learning Representations for All - Xavier Giro-i-Nieto - IRI Barcelona 2020
 
Reservoir computing fast deep learning for sequences
Reservoir computing   fast deep learning for sequencesReservoir computing   fast deep learning for sequences
Reservoir computing fast deep learning for sequences
 

Más de Seldon

TensorFlow London 18: Dr Alastair Moore, Towards the use of Graphical Models ...
TensorFlow London 18: Dr Alastair Moore, Towards the use of Graphical Models ...TensorFlow London 18: Dr Alastair Moore, Towards the use of Graphical Models ...
TensorFlow London 18: Dr Alastair Moore, Towards the use of Graphical Models ...
Seldon
 
TensorFlow London 15: Find bugs in the herd with debuggable TensorFlow code
TensorFlow London 15: Find bugs in the herd with debuggable TensorFlow code TensorFlow London 15: Find bugs in the herd with debuggable TensorFlow code
TensorFlow London 15: Find bugs in the herd with debuggable TensorFlow code
Seldon
 
Tensorflow London 13: Barbara Fusinska 'Hassle Free, Scalable, Machine Learni...
Tensorflow London 13: Barbara Fusinska 'Hassle Free, Scalable, Machine Learni...Tensorflow London 13: Barbara Fusinska 'Hassle Free, Scalable, Machine Learni...
Tensorflow London 13: Barbara Fusinska 'Hassle Free, Scalable, Machine Learni...
Seldon
 
Tensorflow London 13: Zbigniew Wojna 'Deep Learning for Big Scale 2D Imagery'
Tensorflow London 13: Zbigniew Wojna 'Deep Learning for Big Scale 2D Imagery'Tensorflow London 13: Zbigniew Wojna 'Deep Learning for Big Scale 2D Imagery'
Tensorflow London 13: Zbigniew Wojna 'Deep Learning for Big Scale 2D Imagery'
Seldon
 

Más de Seldon (20)

CD4ML and the challenges of testing and quality in ML systems
CD4ML and the challenges of testing and quality in ML systemsCD4ML and the challenges of testing and quality in ML systems
CD4ML and the challenges of testing and quality in ML systems
 
Tensorflow London: Tensorflow and Graph Recommender Networks by Yaz Santissi
Tensorflow London: Tensorflow and Graph Recommender Networks by Yaz SantissiTensorflow London: Tensorflow and Graph Recommender Networks by Yaz Santissi
Tensorflow London: Tensorflow and Graph Recommender Networks by Yaz Santissi
 
TensorFlow London: Progressive Growing of GANs for increased stability, quali...
TensorFlow London: Progressive Growing of GANs for increased stability, quali...TensorFlow London: Progressive Growing of GANs for increased stability, quali...
TensorFlow London: Progressive Growing of GANs for increased stability, quali...
 
TensorFlow London 18: Dr Daniel Martinho-Corbishley, From science to startups...
TensorFlow London 18: Dr Daniel Martinho-Corbishley, From science to startups...TensorFlow London 18: Dr Daniel Martinho-Corbishley, From science to startups...
TensorFlow London 18: Dr Daniel Martinho-Corbishley, From science to startups...
 
TensorFlow London 18: Dr Alastair Moore, Towards the use of Graphical Models ...
TensorFlow London 18: Dr Alastair Moore, Towards the use of Graphical Models ...TensorFlow London 18: Dr Alastair Moore, Towards the use of Graphical Models ...
TensorFlow London 18: Dr Alastair Moore, Towards the use of Graphical Models ...
 
Seldon: Deploying Models at Scale
Seldon: Deploying Models at ScaleSeldon: Deploying Models at Scale
Seldon: Deploying Models at Scale
 
TensorFlow London 17: How NASA Frontier Development Lab scientists use AI to ...
TensorFlow London 17: How NASA Frontier Development Lab scientists use AI to ...TensorFlow London 17: How NASA Frontier Development Lab scientists use AI to ...
TensorFlow London 17: How NASA Frontier Development Lab scientists use AI to ...
 
TensorFlow London 17: Practical Reinforcement Learning with OpenAI
TensorFlow London 17: Practical Reinforcement Learning with OpenAITensorFlow London 17: Practical Reinforcement Learning with OpenAI
TensorFlow London 17: Practical Reinforcement Learning with OpenAI
 
TensorFlow 16: Multimodal Sentiment Analysis with TensorFlow
TensorFlow 16: Multimodal Sentiment Analysis with TensorFlow TensorFlow 16: Multimodal Sentiment Analysis with TensorFlow
TensorFlow 16: Multimodal Sentiment Analysis with TensorFlow
 
TensorFlow 16: Building a Data Science Platform
TensorFlow 16: Building a Data Science Platform TensorFlow 16: Building a Data Science Platform
TensorFlow 16: Building a Data Science Platform
 
Ai in financial services
Ai in financial servicesAi in financial services
Ai in financial services
 
TensorFlow London 15: Find bugs in the herd with debuggable TensorFlow code
TensorFlow London 15: Find bugs in the herd with debuggable TensorFlow code TensorFlow London 15: Find bugs in the herd with debuggable TensorFlow code
TensorFlow London 15: Find bugs in the herd with debuggable TensorFlow code
 
TensorFlow London 14: Ben Hall 'Machine Learning Workloads with Kubernetes an...
TensorFlow London 14: Ben Hall 'Machine Learning Workloads with Kubernetes an...TensorFlow London 14: Ben Hall 'Machine Learning Workloads with Kubernetes an...
TensorFlow London 14: Ben Hall 'Machine Learning Workloads with Kubernetes an...
 
Tensorflow London 13: Barbara Fusinska 'Hassle Free, Scalable, Machine Learni...
Tensorflow London 13: Barbara Fusinska 'Hassle Free, Scalable, Machine Learni...Tensorflow London 13: Barbara Fusinska 'Hassle Free, Scalable, Machine Learni...
Tensorflow London 13: Barbara Fusinska 'Hassle Free, Scalable, Machine Learni...
 
Tensorflow London 13: Zbigniew Wojna 'Deep Learning for Big Scale 2D Imagery'
Tensorflow London 13: Zbigniew Wojna 'Deep Learning for Big Scale 2D Imagery'Tensorflow London 13: Zbigniew Wojna 'Deep Learning for Big Scale 2D Imagery'
Tensorflow London 13: Zbigniew Wojna 'Deep Learning for Big Scale 2D Imagery'
 
TensorFlow London 11: Pierre Harvey Richemond 'Trends and Developments in Rei...
TensorFlow London 11: Pierre Harvey Richemond 'Trends and Developments in Rei...TensorFlow London 11: Pierre Harvey Richemond 'Trends and Developments in Rei...
TensorFlow London 11: Pierre Harvey Richemond 'Trends and Developments in Rei...
 
TensorFlow London 11: Gema Parreno 'Use Cases of TensorFlow'
TensorFlow London 11: Gema Parreno 'Use Cases of TensorFlow'TensorFlow London 11: Gema Parreno 'Use Cases of TensorFlow'
TensorFlow London 11: Gema Parreno 'Use Cases of TensorFlow'
 
Tensorflow London 12: Marcel Horstmann and Laurent Decamp 'Using TensorFlow t...
Tensorflow London 12: Marcel Horstmann and Laurent Decamp 'Using TensorFlow t...Tensorflow London 12: Marcel Horstmann and Laurent Decamp 'Using TensorFlow t...
Tensorflow London 12: Marcel Horstmann and Laurent Decamp 'Using TensorFlow t...
 
TensorFlow London 12: Oliver Gindele 'Recommender systems in Tensorflow'
TensorFlow London 12: Oliver Gindele 'Recommender systems in Tensorflow'TensorFlow London 12: Oliver Gindele 'Recommender systems in Tensorflow'
TensorFlow London 12: Oliver Gindele 'Recommender systems in Tensorflow'
 
TensorFlow London 13.09.17 Ilya Dmitrichenko
TensorFlow London 13.09.17 Ilya DmitrichenkoTensorFlow London 13.09.17 Ilya Dmitrichenko
TensorFlow London 13.09.17 Ilya Dmitrichenko
 

Último

+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
?#DUbAI#??##{{(☎️+971_581248768%)**%*]'#abortion pills for sale in dubai@
 
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers:  A Deep Dive into Serverless Spatial Data and FMECloud Frontiers:  A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
Safe Software
 
Architecting Cloud Native Applications
Architecting Cloud Native ApplicationsArchitecting Cloud Native Applications
Architecting Cloud Native Applications
WSO2
 

Último (20)

MS Copilot expands with MS Graph connectors
MS Copilot expands with MS Graph connectorsMS Copilot expands with MS Graph connectors
MS Copilot expands with MS Graph connectors
 
AXA XL - Insurer Innovation Award Americas 2024
AXA XL - Insurer Innovation Award Americas 2024AXA XL - Insurer Innovation Award Americas 2024
AXA XL - Insurer Innovation Award Americas 2024
 
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost SavingRepurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
Repurposing LNG terminals for Hydrogen Ammonia: Feasibility and Cost Saving
 
Data Cloud, More than a CDP by Matt Robison
Data Cloud, More than a CDP by Matt RobisonData Cloud, More than a CDP by Matt Robison
Data Cloud, More than a CDP by Matt Robison
 
Real Time Object Detection Using Open CV
Real Time Object Detection Using Open CVReal Time Object Detection Using Open CV
Real Time Object Detection Using Open CV
 
ProductAnonymous-April2024-WinProductDiscovery-MelissaKlemke
ProductAnonymous-April2024-WinProductDiscovery-MelissaKlemkeProductAnonymous-April2024-WinProductDiscovery-MelissaKlemke
ProductAnonymous-April2024-WinProductDiscovery-MelissaKlemke
 
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
+971581248768>> SAFE AND ORIGINAL ABORTION PILLS FOR SALE IN DUBAI AND ABUDHA...
 
Apidays Singapore 2024 - Building Digital Trust in a Digital Economy by Veron...
Apidays Singapore 2024 - Building Digital Trust in a Digital Economy by Veron...Apidays Singapore 2024 - Building Digital Trust in a Digital Economy by Veron...
Apidays Singapore 2024 - Building Digital Trust in a Digital Economy by Veron...
 
A Beginners Guide to Building a RAG App Using Open Source Milvus
A Beginners Guide to Building a RAG App Using Open Source MilvusA Beginners Guide to Building a RAG App Using Open Source Milvus
A Beginners Guide to Building a RAG App Using Open Source Milvus
 
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
Web Form Automation for Bonterra Impact Management (fka Social Solutions Apri...
 
Strategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
Strategize a Smooth Tenant-to-tenant Migration and Copilot TakeoffStrategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
Strategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
 
Apidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, Adobe
Apidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, AdobeApidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, Adobe
Apidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, Adobe
 
Apidays New York 2024 - The value of a flexible API Management solution for O...
Apidays New York 2024 - The value of a flexible API Management solution for O...Apidays New York 2024 - The value of a flexible API Management solution for O...
Apidays New York 2024 - The value of a flexible API Management solution for O...
 
EMPOWERMENT TECHNOLOGY GRADE 11 QUARTER 2 REVIEWER
EMPOWERMENT TECHNOLOGY GRADE 11 QUARTER 2 REVIEWEREMPOWERMENT TECHNOLOGY GRADE 11 QUARTER 2 REVIEWER
EMPOWERMENT TECHNOLOGY GRADE 11 QUARTER 2 REVIEWER
 
Corporate and higher education May webinar.pptx
Corporate and higher education May webinar.pptxCorporate and higher education May webinar.pptx
Corporate and higher education May webinar.pptx
 
Polkadot JAM Slides - Token2049 - By Dr. Gavin Wood
Polkadot JAM Slides - Token2049 - By Dr. Gavin WoodPolkadot JAM Slides - Token2049 - By Dr. Gavin Wood
Polkadot JAM Slides - Token2049 - By Dr. Gavin Wood
 
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers:  A Deep Dive into Serverless Spatial Data and FMECloud Frontiers:  A Deep Dive into Serverless Spatial Data and FME
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
 
2024: Domino Containers - The Next Step. News from the Domino Container commu...
2024: Domino Containers - The Next Step. News from the Domino Container commu...2024: Domino Containers - The Next Step. News from the Domino Container commu...
2024: Domino Containers - The Next Step. News from the Domino Container commu...
 
Exploring the Future Potential of AI-Enabled Smartphone Processors
Exploring the Future Potential of AI-Enabled Smartphone ProcessorsExploring the Future Potential of AI-Enabled Smartphone Processors
Exploring the Future Potential of AI-Enabled Smartphone Processors
 
Architecting Cloud Native Applications
Architecting Cloud Native ApplicationsArchitecting Cloud Native Applications
Architecting Cloud Native Applications
 

TensorFlow London: Cutting edge generative models

  • 1. Cutting edge generative models London TensorFlow meetup, March 2019 Pierre Harvey Richemond
  • 2. Table of contents 1. Background 2. New in TensorFlow 2.0 3. Dual use 4. Perspectives 1
  • 4. Background : everything old is new again • Advent of ’modern’ deep learning (resnets, batchnorm) : 2015 (arbitrary depth + large scale training) • If stabilized, deeper is better. • Invention of neural network training rule : 1970 (Linnainmaa’s masters thesis on backpropagation, applied to neural networks) 2
  • 5. Deep vs shallow networks 7x7 conv, 64, /2 pool, /2 3x3 conv, 64 3x3 conv, 64 3x3 conv, 64 3x3 conv, 64 3x3 conv, 64 3x3 conv, 64 3x3 conv, 128, /2 3x3 conv, 128 3x3 conv, 128 3x3 conv, 128 3x3 conv, 128 3x3 conv, 128 3x3 conv, 128 3x3 conv, 128 3x3 conv, 256, /2 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 512, /2 3x3 conv, 512 3x3 conv, 512 3x3 conv, 512 3x3 conv, 512 3x3 conv, 512 avg pool fc 1000 image 3x3 conv, 512 3x3 conv, 64 3x3 conv, 64 pool, /2 3x3 conv, 128 3x3 conv, 128 pool, /2 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 pool, /2 3x3 conv, 512 3x3 conv, 512 3x3 conv, 512 pool, /2 3x3 conv, 512 3x3 conv, 512 3x3 conv, 512 3x3 conv, 512 pool, /2 fc 4096 fc 4096 fc 1000 image output size: 112 output size: 224 output size: 56 output size: 28 output size: 14 output size: 7 output size: 1 VGG-19 34-layer plain 7x7 conv, 64, /2 pool, /2 3x3 conv, 64 3x3 conv, 64 3x3 conv, 64 3x3 conv, 64 3x3 conv, 64 3x3 conv, 64 3x3 conv, 128, /2 3x3 conv, 128 3x3 conv, 128 3x3 conv, 128 3x3 conv, 128 3x3 conv, 128 3x3 conv, 128 3x3 conv, 128 3x3 conv, 256, /2 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 256 3x3 conv, 512, /2 3x3 conv, 512 3x3 conv, 512 3x3 conv, 512 3x3 conv, 512 3x3 conv, 512 avg pool fc 1000 image 34-layer residual 3
  • 6. Why does it work now ? • In between : GPUs, fibre broadband, handheld megapixel cameras, automatic differentiation software. • All contributed to the explosion of applications we see today. • Deep networks tend to work better and automatically find good features for data classification & explanation on natural data: images/video and sound. • Other domains like text, structured data (chemistry...) can work too with careful network design. 4
  • 7. Deep Learning breakthroughs - 2014 • VGG architecture • DeepDream • First formulation of GANs • First formulation of VAEs • Black-box Variational Inference • Neural Turing Machines • Attention mechanisms 5
  • 8. Deep Learning breakthroughs - 2015 • Residual Networks • Batch Normalization • ADAM optimizer • ELU activation function • Neural style transfer • Graph convolutional networks • Visual question answering (Karpathy) • Normalizing Flows • Deep Q-learning on Atari games • Keras, TensorFlow 6
  • 9. Deep Learning breakthroughs - 2016 • Wide Residual Networks • DenseNets • WaveNet • PixelCNN • Asynchronous Advantage Actor Critic (A3C) • AlphaGo • Neural Machine Translation at scale • PyTorch 7
  • 10. Deep Learning breakthroughs - 2017 • Wasserstein GAN • Progressive Growing of GANs • Capsule Networks • Restarted SGD - cyclical learning rates • AlphaGo Zero (MCTS as policy improvement ; no human knowledge involved) • Distributional Reinforcement Learning • Equivalence of softmax Q-learning and entropic policy gradients • TensorFlow distributions, Pyro 8
  • 11. Deep Learning breakthroughs - 2018 • Spectral Normalization for GANs • Fast initialization of convolutional neural networks • Deep Video Portraits • Differentiable neural architecture search • OpenAI Five • Video-to-video synthesis • StyleGAN 9
  • 12. A success story : Generative models • Generative modelling aims at automatically explain features, and generate new instances, of datapoints in a dataset • Two main approaches that both use 2 different neural networks instead of 1. • Applications of the field to deep learning are 4 years old • A success story of engineering 10
  • 13. Generative models : GANs Figure 1: Generative Adversarial Networks (GANs). A generator (counterfeinter) and a discriminator (police) network play an adversarial game, whose Nash equilibrium is perfect replication of the data distribution. 11
  • 14. Generative models : VAEs Figure 2: A typical variational autoencoder (VAE) architecture. An encoder pushes input data through a (low) dimensionality bottleneck, that learns relevant features in a latent code (optimized probabilistically). The decoder network attempts to reconstruct inputs. 12
  • 15. Illustrations - 2015 Figure 3: MNIST digits toy dataset. Interpolation. 13
  • 16. Illustrations - 2016 Figure 4: Interpolations varying the strength of a smile vector computed by doing latent space averaging of labeled pictures. Courtesy of Tom White. 14
  • 17. Illustrations - 2017 Figure 5: GAN-generated 1024*1024 portraits from Karras et al., 2017. 15
  • 18. Progressive Growing of GANs - Results, late 2018 Figure 6: High-resolution generations from Karras et al., 2018. 16
  • 20. TensorFlow 2.0 - General layout Figure 7: Eager (dynamic graphing) and more Pythonic. 17
  • 21. TensorFlow 2.0 - Layers declaration Figure 8: Object-oriented layers. 18
  • 22. TensorFlow 2.0 - Functional Figure 9: The ’functional’ API, from Keras, is available. 19
  • 23. TensorFlow 2.0 - VAE example Figure 10: The functional API can shorten code significantly ! 20
  • 25. From fake news to un-real people www.thispersondoesnotexist.com A blog post by Kyle McDonald on how to detect fake portraits: 21
  • 26. Some photorealistic transformations - 2018 Deep Video Portraits (Portrait reenactment) Vid2Vid (Photorealistic video-to-video translation) See also TacoTron for realistic speech generation. 22
  • 27. OpenAI GPT-2 - Principle • A generative model trained at scale • A relatively ’simple’ architecture and prediction concept, but the scaling makes all the difference (8 million webpages training set) • While it can be used in a multitask context ( question answering, text summarization, translation...), it really shines in text generation. • For the first time, OpenAI has declined releasing the weights of the full trained model, citing risks • This has caused an unprecedented stir in the (self-regulated) machine learning research community 23
  • 28. OpenAI GPT-2 - Coherent paragraph generation - 1 When initalized with the following prompt text : ’In a shocking finding, scientist discovered a herd of unicorns living in a remote, previously unexplored valley, in the Andes Mountains. Even more surprising to the researchers was the fact that the unicorns spoke perfect English.’ the model generated as a follow-up... 24
  • 29. OpenAI GPT-2 - Coherent paragraph generation - 2 ’ “The scientist named the population, after their distinctive horn, Ovid’s Unicorn. These four-horned, silver-white unicorns were previously unknown to science. Now, after almost two centuries, the mystery of what sparked this odd phenomenon is finally solved. Dr. Jorge Pérez, an evolutionary biologist from the University of La Paz, and several companions, were exploring the Andes Mountains when they found a small valley, with no other animals or humans. Pérez noticed that the valley had what appeared to be a natural fountain, surrounded by two peaks of rock and silver snow. Pérez and the others then ventured further into the valley. ‘By the time we reached the top of one peak, the water looked blue, with some crystals on top,’ said Pérez. Pérez and his friends were astonished to see the unicorn herd. These creatures could be seen from the air without having to move too much to see them – they were so close they could touch their horns.”’ 25
  • 30. OpenAI GPT-2 : Test 1 - Neural news generation GPT2 seed ’DeepMind should have been a UK champion says ex-Google CFO who took it to America’: Google executives Mark Zuckerberg and Sergey Brin,formerly head of the social media giant, have been accused of lying to the US government about the extent of their ties to the Kremlin. 26
  • 31. OpenAI GPT-2 : Test 2 - Scientific Generation ! ’We compute a stochastic scaling function for data with a weighted average of the mean as the value for which to apply an Likert distribution over the data in the gradient descent. We then evaluate the prediction accuracy and find that the weights are not completely uniform in this way. Instead, the weight of this residual model is a weighted average of the data in the gradient descent on the average, which is also very accurate.’ 27
  • 32. Towards post-truth economics • Shattering the realism barrier of the uncanny valley • Legislation will be required (don’t hold your breath) • Content industries could be notably disrupted • Is this a serious use case for blockchain tech ? Is the marginal cost of content creation going to zero ? 28
  • 34. Are we at AGI yet ? 3 different views : 1. Killer robots are coming, let’s ban research ! 2. AI is just matrix multiplication... 3. AI/Deep Learning is in fact letting an optimization algorithm automatically write code for you 29
  • 35. ’Science sans conscience’ : TBC... Figure 11: Human DNA data embedding and clustering (UMAP), categorized by ethnicity. BiorXiv 2018, ’Revealing multi-scale population structure in large cohorts’. 30
  • 37. Thank you for your attention Twitter @KloudStrife www.deeplearningmathematics.com