SlideShare a Scribd company logo
1 of 59
Download to read offline
Deep Learning:
State of the Art (2019)
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: State of the Art*
(Breakthrough Developments in 2017 & 2018)
• BERT and Natural Language Processing
• Tesla Autopilot Hardware v2+: Neural Networks at Scale
• AdaNet: AutoML with Ensembles
• AutoAugment: Deep RL Data Augmentation
• Training Deep Networks with Synthetic Data
• Segmentation Annotation with Polygon-RNN++
• DAWNBench: Training Fast and Cheap
• BigGAN: State of the Art in Image Synthesis
• Video-to-Video Synthesis
• Semantic Segmentation
• AlphaZero & OpenAI Five
• Deep Learning Frameworks
* This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an
overview of exciting recent developments.
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Encoder-Decoder Architecture
Sequence-to-Sequence Model - Neural Machine Translation
[295]
Encoder RNN encodes input sequence into a fixed size vector,
and then is passed repeatedly to decoder RNN.
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Attention
[296]
Attention mechanism allows the network to refer back to the
input sequence, instead of forcing it to encode all information
into one fixed-length vector.
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Attention
[312]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Attention
[312]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Self-Attention
[312]
"The animal didn't cross the street because it was too tired"
More details: http://jalammar.github.io/illustrated-transformer/
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Transformer
[312]
Vaswani, Ashish, et al. "Attention is all you need." Advances in Neural Information Processing Systems. 2017.
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
Word Embeddings
(Word2Vec)
[297]
Word Vector
Skip Gram Model:
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Context-Aware Embeddings
[312]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
OpenAI Transformer
1. Pre-train a Transformer’s decoder for language modeling
2. Train it on, for example, a sentence classification task
[312]
1 2
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
BERT
Devlin, Jacob, et al. "Bert: Pre-training of deep bidirectional transformers for language understanding." (2018).
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
BERT Applications
Now you can use BERT:
• Create contextualized word
embeddings (like ELMo)
• Sentence classification
• Sentence pair classification
• Sentence pair similarity
• Multiple choice
• Sentence tagging
• Question answering
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: State of the Art*
(Breakthrough Developments in 2017 & 2018)
• BERT and Natural Language Processing
• Tesla Autopilot Hardware v2+: Neural Networks at Scale
• AdaNet: AutoML with Ensembles
• AutoAugment: Deep RL Data Augmentation
• Training Deep Networks with Synthetic Data
• Segmentation Annotation with Polygon-RNN++
• DAWNBench: Training Fast and Cheap
• BigGAN: State of the Art in Image Synthesis
• Video-to-Video Synthesis
• Semantic Segmentation
• AlphaZero & OpenAI Five
• Deep Learning Frameworks
* This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an
overview of exciting recent developments.
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Tesla Autopilot Hardware v2+
• Specialized NVIDIA Drive PX 2 hardware
• Neural network takes all 8 cameras as input
• Based on Inception v1 architecture
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Autopilot Reaches 1 Billion Miles
(~0.5 Billion on Hardware v2+)
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: State of the Art*
(Breakthrough Developments in 2017 & 2018)
• BERT and Natural Language Processing
• Tesla Autopilot Hardware v2+: Neural Networks at Scale
• AdaNet: AutoML with Ensembles
• AutoAugment: Deep RL Data Augmentation
• Training Deep Networks with Synthetic Data
• Segmentation Annotation with Polygon-RNN++
• DAWNBench: Training Fast and Cheap
• BigGAN: State of the Art in Image Synthesis
• Video-to-Video Synthesis
• Semantic Segmentation
• AlphaZero & OpenAI Five
• Deep Learning Frameworks
* This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an
overview of exciting recent developments.
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
AutoML and Neural Architecture Search (NASNet)
[300, 301]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
AdaNet
[317]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: State of the Art*
(Breakthrough Developments in 2017 & 2018)
• BERT and Natural Language Processing
• Tesla Autopilot Hardware v2+: Neural Networks at Scale
• AdaNet: AutoML with Ensembles
• AutoAugment: Deep RL Data Augmentation
• Training Deep Networks with Synthetic Data
• Segmentation Annotation with Polygon-RNN++
• DAWNBench: Training Fast and Cheap
• BigGAN: State of the Art in Image Synthesis
• Video-to-Video Synthesis
• Semantic Segmentation
• AlphaZero & OpenAI Five
• Deep Learning Frameworks
* This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an
overview of exciting recent developments.
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
AutoAugment: RL for Data Augmentation
[318]
• Show that transfer learning can also be done augmentation
policies instead of weights (or with).
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: State of the Art*
(Breakthrough Developments in 2017 & 2018)
• BERT and Natural Language Processing
• Tesla Autopilot Hardware v2+: Neural Networks at Scale
• AdaNet: AutoML with Ensembles
• AutoAugment: Deep RL Data Augmentation
• Training Deep Networks with Synthetic Data
• Segmentation Annotation with Polygon-RNN++
• DAWNBench: Training Fast and Cheap
• BigGAN: State of the Art in Image Synthesis
• Video-to-Video Synthesis
• Semantic Segmentation
• AlphaZero & OpenAI Five
• Deep Learning Frameworks
* This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an
overview of exciting recent developments.
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Training on Randomized Synthetic Data
[319]
• number and types of objects
• number, types, colors, and scales of distractors
• texture on the object of interest, and background photograph
• location of the virtual camera with respect to the scene
• angle of the camera with respect to the scene
• number and locations of point lights
Tremblay, Jonathan, et al. "Training deep networks with synthetic data:
Bridging the reality gap by domain randomization." (2018).
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: State of the Art*
(Breakthrough Developments in 2017 & 2018)
• BERT and Natural Language Processing
• Tesla Autopilot Hardware v2+: Neural Networks at Scale
• AdaNet: AutoML with Ensembles
• AutoAugment: Deep RL Data Augmentation
• Training Deep Networks with Synthetic Data
• Segmentation Annotation with Polygon-RNN++
• DAWNBench: Training Fast and Cheap
• BigGAN: State of the Art in Image Synthesis
• Video-to-Video Synthesis
• Semantic Segmentation
• AlphaZero & OpenAI Five
• Deep Learning Frameworks
* This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an
overview of exciting recent developments.
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Segmentation Annotation with Polygon-RNN++
• High-resolution polygon
segmentation generation
with human-in-the-loop
[319]
Acuna, David, et al. "Efficient Interactive Annotation of
Segmentation Datasets With Polygon-RNN++." CVPR 2018.
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: State of the Art*
(Breakthrough Developments in 2017 & 2018)
• BERT and Natural Language Processing
• Tesla Autopilot Hardware v2+: Neural Networks at Scale
• AdaNet: AutoML with Ensembles
• AutoAugment: Deep RL Data Augmentation
• Training Deep Networks with Synthetic Data
• Segmentation Annotation with Polygon-RNN++
• DAWNBench: Training Fast and Cheap
• BigGAN: State of the Art in Image Synthesis
• Video-to-Video Synthesis
• Semantic Segmentation
• AlphaZero & OpenAI Five
• Deep Learning Frameworks
* This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an
overview of exciting recent developments.
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
DAWNBench: Training Fast and Cheap
DAWNBench:
• Competition on speed and cost of
training and inference that achieves
93% for ImageNet and 94% for CIFAR
10
fast.ai – Training:
• ImageNet in 3 hours for $25
• CIFAR10 for $0.26
• Key idea: During training, if you very
slowly increase learning rate while
decreasing momentum, you can
train at extremely high learning
rates, thus avoiding over-fitting, and
training in far fewer epochs.
• Details: http://bit.ly/2H6yv6H
[320]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: State of the Art*
(Breakthrough Developments in 2017 & 2018)
• BERT and Natural Language Processing
• Tesla Autopilot Hardware v2+: Neural Networks at Scale
• AdaNet: AutoML with Ensembles
• AutoAugment: Deep RL Data Augmentation
• Training Deep Networks with Synthetic Data
• Segmentation Annotation with Polygon-RNN++
• DAWNBench: Training Fast and Cheap
• BigGAN: State of the Art in Image Synthesis
• Video-to-Video Synthesis
• Semantic Segmentation
• AlphaZero & OpenAI Five
• Deep Learning Frameworks
* This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an
overview of exciting recent developments.
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
BigGAN: State of the Art in Image Synthesis
• Same GAN techniques, much larger scale
• Increase model capacity + increase batch size
[321]
Brock, Andrew, Jeff Donahue, and Karen Simonyan. "Large scale gan
training for high fidelity natural image synthesis." (2018).
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: State of the Art*
(Breakthrough Developments in 2017 & 2018)
• BERT and Natural Language Processing
• Tesla Autopilot Hardware v2+: Neural Networks at Scale
• AdaNet: AutoML with Ensembles
• AutoAugment: Deep RL Data Augmentation
• Training Deep Networks with Synthetic Data
• Segmentation Annotation with Polygon-RNN++
• DAWNBench: Training Fast and Cheap
• BigGAN: State of the Art in Image Synthesis
• Video-to-Video Synthesis
• Semantic Segmentation
• AlphaZero & OpenAI Five
• Deep Learning Frameworks
* This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an
overview of exciting recent developments.
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Video-to-Video Synthesis
• Model temporal dynamics by adding a spatio-temporal
adversarial objective to the GAN framework
[322]
Wang, Ting-Chun, et al. "Video-to-video synthesis." (2018).
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Video-to-Video Synthesis
• Model temporal dynamics by adding a spatio-temporal
adversarial objective to the GAN framework
[322]
Wang, Ting-Chun, et al. "Video-to-video synthesis." (2018).
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Video-to-Video Synthesis
• Model temporal dynamics by adding a spatio-temporal
adversarial objective to the GAN framework
[322]
Wang, Ting-Chun, et al. "Video-to-video synthesis." (2018).
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: State of the Art*
(Breakthrough Developments in 2017 & 2018)
• BERT and Natural Language Processing
• Tesla Autopilot Hardware v2+: Neural Networks at Scale
• AdaNet: AutoML with Ensembles
• AutoAugment: Deep RL Data Augmentation
• Training Deep Networks with Synthetic Data
• Segmentation Annotation with Polygon-RNN++
• DAWNBench: Training Fast and Cheap
• BigGAN: State of the Art in Image Synthesis
• Video-to-Video Synthesis
• Semantic Segmentation
• AlphaZero & OpenAI Five
• Deep Learning Frameworks
* This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an
overview of exciting recent developments.
2019https://deeplearning.mit.edu
• AlexNet (2012): First CNN (15.4%)
• 8 layers
• 61 million parameters
• ZFNet (2013): 15.4% to 11.2%
• 8 layers
• More filters. Denser stride.
• VGGNet (2014): 11.2% to 7.3%
• Beautifully uniform:
3x3 conv, stride 1, pad 1, 2x2 max pool
• 16 layers
• 138 million parameters
• GoogLeNet (2014): 11.2% to 6.7%
• Inception modules
• 22 layers
• 5 million parameters
(throw away fully connected layers)
• ResNet (2015): 6.7% to 3.57%
• More layers = better performance
• 152 layers
• CUImage (2016): 3.57% to 2.99%
• Ensemble of 6 models
• SENet (2017): 2.99% to 2.251%
• Squeeze and excitation block: network
is allowed to adaptively adjust the
weighting of each feature map in the
convolutional block.
References: [90]
Human error (5.1%)
surpassed in 2015
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Object Detection / Localization
Region-Based Methods | Shown: Faster R-CNN
[299]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Object Detection / Localization
Single-Shot Methods | Shown: SSD
[299]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Object Detection: State of the Art Progress
[326]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Semantic Segmentation
[175]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
State-of-the-Art: DeepLab v3
• PASCAL VOC Challenge: http://bit.ly/2HdzTEu
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
FCN (Nov 2014)
Paper: “Fully Convolutional Networks for Semantic Segmentation”
• Repurpose Imagenet pretrained nets
• Upsample using deconvolution
• Skip connections to improve coarseness of upsampling
[173]
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
SegNet (Nov 2015)
Paper: “SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation”
• Maxpooling indices transferred to decoder to improve the
segmentation resolution.
[175]
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
Dilated Convolutions (Nov 2015)
Paper: “Multi-Scale Context Aggregation by Dilated Convolutions”
• Since pooling decreases resolution:
• Added “dilated convolution layer”
• Still interpolate up from 1/8 of original image size
[175]
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
DeepLab v1, v2 (Jun 2016)
Paper: “DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and
Fully Connected CRFs”
• Added fully-connected Conditional Random Fields (CRFs) – as a
post-processing step
• Smooth segmentation based on the underlying image intensities
[179]
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
Key Aspects of Segmentation
• Fully convolutional networks (FCNs) - replace fully-connected
layers with convolutional layers. Classification network is where
the biggest gains come from.
• Deeper, updated models (ResNet, etc) consistent with ImageNet
Challenge object classification tasks.
• Conditional Random Fields (CRFs) to capture both local and
long-range dependencies within an image to refine the
prediction map.
• Dilated convolution (aka Atrous convolution) – maintain
computational cost, increase resolution of intermediate feature
maps
• Process at multiple scales and combine the information
together
[174, 176]
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
DeepLab v3
Paper: “Rethinking Atrous Convolution for Semantic Image Segmentation”
• Multi-scale processing, without increasing parameters.
• Increasing “atrous rate” enlarges the model’s “field-of-view”
[179]
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
DeepLab v3 trained on CityScapes
Tutorial: https://github.com/lexfridman/mit-deep-learning
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: State of the Art*
(Breakthrough Developments in 2017 & 2018)
• BERT and Natural Language Processing
• Tesla Autopilot Hardware v2+: Neural Networks at Scale
• AdaNet: AutoML with Ensembles
• AutoAugment: Deep RL Data Augmentation
• Training Deep Networks with Synthetic Data
• Segmentation Annotation with Polygon-RNN++
• DAWNBench: Training Fast and Cheap
• BigGAN: State of the Art in Image Synthesis
• Video-to-Video Synthesis
• Semantic Segmentation
• AlphaZero & OpenAI Five
• Deep Learning Frameworks
* This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an
overview of exciting recent developments.
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
Atari Breakout
[85]
After
120 Minutes
of Training
After
10 Minutes
of Training
After
240 Minutes
of Training
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
DQN: Atari Games (2015)
[83]
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
AlphaGo (2016): Beat Top Human at Go
[83]
2019https://deeplearning.mit.eduFor the full updated list of references visit:
https://selfdrivingcars.mit.edu/references
AlphaGo Zero (2017): Beats AlphaGo
[149]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references [324]
AlphaZero (Dec 2017) vs StockFish (Chess) & Elmo (Shogi)
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
OpenAI & Dota 2
• Dota 2 as a testbed for the messiness and continuous nature of
the real world: teamwork, long time horizons, and hidden
information.
[325]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
OpenAI & Dota 2 Progress
• Aug, 2017: 1v1 bot beats top professional Dota 2 players.
• Aug, 2018: OpenAI Five lost two games against top Dota 2 players at The
International. “We are looking forward to pushing Five to the next level.”
[325]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: State of the Art*
(Breakthrough Developments in 2017 & 2018)
• BERT and Natural Language Processing
• Tesla Autopilot Hardware v2+: Neural Networks at Scale
• AdaNet: AutoML with Ensembles
• AutoAugment: Deep RL Data Augmentation
• Training Deep Networks with Synthetic Data
• Segmentation Annotation with Polygon-RNN++
• DAWNBench: Training Fast and Cheap
• BigGAN: State of the Art in Image Synthesis
• Video-to-Video Synthesis
• Semantic Segmentation
• AlphaZero & OpenAI Five
• Deep Learning Frameworks
* This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an
overview of exciting recent developments.
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning Frameworks
[327]
Deep Learning Framework
Power Scores (by Jeff Hale)
http://bit.ly/2GBa3tU
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
Factors to consider:
• Learning curve
• Speed of development
• Size and passion of community
• Number of papers implemented in framework
• Likelihood of long-term growth and stability
• Ecosystem of tooling
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Deep Learning: 2019 and Beyond
• On backpropagation:
“My view is throw it all
away and start again.”
• “The future depends on
some graduate student
who is deeply suspicious
of everything I have said.”
- Geoffrey Hinton
“Godfather of Deep Learning”
[328]
2019https://deeplearning.mit.eduFor the full list of references visit:
https://hcai.mit.edu/references
Thank You
Website:
deeplearning.mit.edu
• Videos and slides posted on the website
• Code posted on GitHub:
https://github.com/lexfridman/mit-deep-learning

More Related Content

What's hot

Predicting Influence and Communities Using Graph Algorithms
Predicting Influence and Communities Using Graph AlgorithmsPredicting Influence and Communities Using Graph Algorithms
Predicting Influence and Communities Using Graph AlgorithmsDatabricks
 
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 LibraryTigerGraph
 
Graph Data Science at Scale
Graph Data Science at ScaleGraph Data Science at Scale
Graph Data Science at ScaleNeo4j
 
PFHub: Phase Field Community Hub
PFHub: Phase Field Community HubPFHub: Phase Field Community Hub
PFHub: Phase Field Community HubDaniel Wheeler
 
The LINQ Between XML and Database
The LINQ Between XML and DatabaseThe LINQ Between XML and Database
The LINQ Between XML and DatabaseIRJET Journal
 
Deep Learning for Autonomous Driving
Deep Learning for Autonomous DrivingDeep Learning for Autonomous Driving
Deep Learning for Autonomous DrivingJan Wiegelmann
 
Elegant and Scalable Code Querying with Code Property Graphs
Elegant and Scalable Code Querying with Code Property GraphsElegant and Scalable Code Querying with Code Property Graphs
Elegant and Scalable Code Querying with Code Property GraphsConnected Data World
 

What's hot (10)

Predicting Influence and Communities Using Graph Algorithms
Predicting Influence and Communities Using Graph AlgorithmsPredicting Influence and Communities Using Graph Algorithms
Predicting Influence and Communities Using Graph Algorithms
 
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
 
ID 259 Poster
ID 259 PosterID 259 Poster
ID 259 Poster
 
Graph Data Science at Scale
Graph Data Science at ScaleGraph Data Science at Scale
Graph Data Science at Scale
 
PFHub: Phase Field Community Hub
PFHub: Phase Field Community HubPFHub: Phase Field Community Hub
PFHub: Phase Field Community Hub
 
The LINQ Between XML and Database
The LINQ Between XML and DatabaseThe LINQ Between XML and Database
The LINQ Between XML and Database
 
Deep Learning for Autonomous Driving
Deep Learning for Autonomous DrivingDeep Learning for Autonomous Driving
Deep Learning for Autonomous Driving
 
CV
CVCV
CV
 
Elegant and Scalable Code Querying with Code Property Graphs
Elegant and Scalable Code Querying with Code Property GraphsElegant and Scalable Code Querying with Code Property Graphs
Elegant and Scalable Code Querying with Code Property Graphs
 
Semantics2014
Semantics2014Semantics2014
Semantics2014
 

Similar to Deep learning state_of_the_art- Autonomous Driving

Introduction to Nebula Graph, an Open-Source Distributed Graph Database
Introduction to Nebula Graph, an Open-Source Distributed Graph DatabaseIntroduction to Nebula Graph, an Open-Source Distributed Graph Database
Introduction to Nebula Graph, an Open-Source Distributed Graph DatabaseNebula Graph
 
The Data Platform Administration Handling the 100 PB.pdf
The Data Platform Administration Handling the 100 PB.pdfThe Data Platform Administration Handling the 100 PB.pdf
The Data Platform Administration Handling the 100 PB.pdfRakuten Group, Inc.
 
Enabling Data centric Teams
Enabling Data centric TeamsEnabling Data centric Teams
Enabling Data centric TeamsData Con LA
 
Deep Learning for Recommender Systems
Deep Learning for Recommender SystemsDeep Learning for Recommender Systems
Deep Learning for Recommender SystemsNick Pentreath
 
Austin,TX Meetup presentation tensorflow final oct 26 2017
Austin,TX Meetup presentation tensorflow final oct 26 2017Austin,TX Meetup presentation tensorflow final oct 26 2017
Austin,TX Meetup presentation tensorflow final oct 26 2017Clarisse Hedglin
 
Neo4j – The Fastest Path to Scalable Real-Time Analytics
Neo4j – The Fastest Path to Scalable Real-Time AnalyticsNeo4j – The Fastest Path to Scalable Real-Time Analytics
Neo4j – The Fastest Path to Scalable Real-Time AnalyticsNeo4j
 
Better Hackathon 2020 - Fraunhofer IAIS - Semantic geo-clustering with SANSA
Better Hackathon 2020 - Fraunhofer IAIS - Semantic geo-clustering with SANSABetter Hackathon 2020 - Fraunhofer IAIS - Semantic geo-clustering with SANSA
Better Hackathon 2020 - Fraunhofer IAIS - Semantic geo-clustering with SANSAPRBETTER
 
Nitin_updated_Profile
Nitin_updated_ProfileNitin_updated_Profile
Nitin_updated_ProfileNitin Saxena
 
Parallel universe-issue-29
Parallel universe-issue-29Parallel universe-issue-29
Parallel universe-issue-29DESMOND YUEN
 
INTERFACE, by apidays - The Evolution of Data Movement.pdf
INTERFACE, by apidays - The Evolution of Data Movement.pdfINTERFACE, by apidays - The Evolution of Data Movement.pdf
INTERFACE, by apidays - The Evolution of Data Movement.pdfapidays
 
MIT Deep Learning Basics: Introduction and Overview by Lex Fridman
MIT Deep Learning Basics: Introduction and Overview by Lex FridmanMIT Deep Learning Basics: Introduction and Overview by Lex Fridman
MIT Deep Learning Basics: Introduction and Overview by Lex FridmanPeerasak C.
 
Serverless machine learning architectures at Helixa
Serverless machine learning architectures at HelixaServerless machine learning architectures at Helixa
Serverless machine learning architectures at HelixaData Science Milan
 
Still on IBM BigInsights? We have the right path for you
Still on IBM BigInsights? We have the right path for youStill on IBM BigInsights? We have the right path for you
Still on IBM BigInsights? We have the right path for youModusOptimum
 
Compositional AI: Fusion of AI/ML Services
Compositional AI: Fusion of AI/ML ServicesCompositional AI: Fusion of AI/ML Services
Compositional AI: Fusion of AI/ML ServicesDebmalya Biswas
 
Gimel and PayPal Notebooks @ TDWI Leadership Summit Orlando
Gimel and PayPal Notebooks @ TDWI Leadership Summit OrlandoGimel and PayPal Notebooks @ TDWI Leadership Summit Orlando
Gimel and PayPal Notebooks @ TDWI Leadership Summit OrlandoRomit Mehta
 
ArangoML Pipeline Cloud - Managed Machine Learning Metadata
ArangoML Pipeline Cloud - Managed Machine Learning MetadataArangoML Pipeline Cloud - Managed Machine Learning Metadata
ArangoML Pipeline Cloud - Managed Machine Learning MetadataArangoDB Database
 
Inteligencia artificial, open source e IBM Call for Code
Inteligencia artificial, open source e IBM Call for CodeInteligencia artificial, open source e IBM Call for Code
Inteligencia artificial, open source e IBM Call for CodeLuciano Resende
 
Future of jobs and digital economy citi conference 090618
Future of jobs and digital economy citi conference 090618Future of jobs and digital economy citi conference 090618
Future of jobs and digital economy citi conference 090618Economic Strategy Institute
 
Nodes 2023 - Knowledge graph based chatbot.pptx
Nodes 2023 - Knowledge graph based chatbot.pptxNodes 2023 - Knowledge graph based chatbot.pptx
Nodes 2023 - Knowledge graph based chatbot.pptxTomazBratanic1
 

Similar to Deep learning state_of_the_art- Autonomous Driving (20)

Introduction to Nebula Graph, an Open-Source Distributed Graph Database
Introduction to Nebula Graph, an Open-Source Distributed Graph DatabaseIntroduction to Nebula Graph, an Open-Source Distributed Graph Database
Introduction to Nebula Graph, an Open-Source Distributed Graph Database
 
The Data Platform Administration Handling the 100 PB.pdf
The Data Platform Administration Handling the 100 PB.pdfThe Data Platform Administration Handling the 100 PB.pdf
The Data Platform Administration Handling the 100 PB.pdf
 
Enabling Data centric Teams
Enabling Data centric TeamsEnabling Data centric Teams
Enabling Data centric Teams
 
Deep Learning for Recommender Systems
Deep Learning for Recommender SystemsDeep Learning for Recommender Systems
Deep Learning for Recommender Systems
 
Austin,TX Meetup presentation tensorflow final oct 26 2017
Austin,TX Meetup presentation tensorflow final oct 26 2017Austin,TX Meetup presentation tensorflow final oct 26 2017
Austin,TX Meetup presentation tensorflow final oct 26 2017
 
Neo4j – The Fastest Path to Scalable Real-Time Analytics
Neo4j – The Fastest Path to Scalable Real-Time AnalyticsNeo4j – The Fastest Path to Scalable Real-Time Analytics
Neo4j – The Fastest Path to Scalable Real-Time Analytics
 
Better Hackathon 2020 - Fraunhofer IAIS - Semantic geo-clustering with SANSA
Better Hackathon 2020 - Fraunhofer IAIS - Semantic geo-clustering with SANSABetter Hackathon 2020 - Fraunhofer IAIS - Semantic geo-clustering with SANSA
Better Hackathon 2020 - Fraunhofer IAIS - Semantic geo-clustering with SANSA
 
Nitin_updated_Profile
Nitin_updated_ProfileNitin_updated_Profile
Nitin_updated_Profile
 
Parallel universe-issue-29
Parallel universe-issue-29Parallel universe-issue-29
Parallel universe-issue-29
 
INTERFACE, by apidays - The Evolution of Data Movement.pdf
INTERFACE, by apidays - The Evolution of Data Movement.pdfINTERFACE, by apidays - The Evolution of Data Movement.pdf
INTERFACE, by apidays - The Evolution of Data Movement.pdf
 
MIT Deep Learning Basics: Introduction and Overview by Lex Fridman
MIT Deep Learning Basics: Introduction and Overview by Lex FridmanMIT Deep Learning Basics: Introduction and Overview by Lex Fridman
MIT Deep Learning Basics: Introduction and Overview by Lex Fridman
 
Serverless machine learning architectures at Helixa
Serverless machine learning architectures at HelixaServerless machine learning architectures at Helixa
Serverless machine learning architectures at Helixa
 
Still on IBM BigInsights? We have the right path for you
Still on IBM BigInsights? We have the right path for youStill on IBM BigInsights? We have the right path for you
Still on IBM BigInsights? We have the right path for you
 
Compositional AI: Fusion of AI/ML Services
Compositional AI: Fusion of AI/ML ServicesCompositional AI: Fusion of AI/ML Services
Compositional AI: Fusion of AI/ML Services
 
Gimel and PayPal Notebooks @ TDWI Leadership Summit Orlando
Gimel and PayPal Notebooks @ TDWI Leadership Summit OrlandoGimel and PayPal Notebooks @ TDWI Leadership Summit Orlando
Gimel and PayPal Notebooks @ TDWI Leadership Summit Orlando
 
ArangoML Pipeline Cloud - Managed Machine Learning Metadata
ArangoML Pipeline Cloud - Managed Machine Learning MetadataArangoML Pipeline Cloud - Managed Machine Learning Metadata
ArangoML Pipeline Cloud - Managed Machine Learning Metadata
 
Ibm db2 big sql
Ibm db2 big sqlIbm db2 big sql
Ibm db2 big sql
 
Inteligencia artificial, open source e IBM Call for Code
Inteligencia artificial, open source e IBM Call for CodeInteligencia artificial, open source e IBM Call for Code
Inteligencia artificial, open source e IBM Call for Code
 
Future of jobs and digital economy citi conference 090618
Future of jobs and digital economy citi conference 090618Future of jobs and digital economy citi conference 090618
Future of jobs and digital economy citi conference 090618
 
Nodes 2023 - Knowledge graph based chatbot.pptx
Nodes 2023 - Knowledge graph based chatbot.pptxNodes 2023 - Knowledge graph based chatbot.pptx
Nodes 2023 - Knowledge graph based chatbot.pptx
 

Recently uploaded

Somya Surve Escorts Service Bilaspur ❣️ 7014168258 ❣️ High Cost Unlimited Har...
Somya Surve Escorts Service Bilaspur ❣️ 7014168258 ❣️ High Cost Unlimited Har...Somya Surve Escorts Service Bilaspur ❣️ 7014168258 ❣️ High Cost Unlimited Har...
Somya Surve Escorts Service Bilaspur ❣️ 7014168258 ❣️ High Cost Unlimited Har...nirzagarg
 
FULL NIGHT — 9999894380 Call Girls In Jagat Puri | Delhi
FULL NIGHT — 9999894380 Call Girls In Jagat Puri | DelhiFULL NIGHT — 9999894380 Call Girls In Jagat Puri | Delhi
FULL NIGHT — 9999894380 Call Girls In Jagat Puri | DelhiSaketCallGirlsCallUs
 
Majestic Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...
Majestic Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...Majestic Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...
Majestic Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...amitlee9823
 
如何办理麦考瑞大学毕业证(MQU毕业证书)成绩单原版一比一
如何办理麦考瑞大学毕业证(MQU毕业证书)成绩单原版一比一如何办理麦考瑞大学毕业证(MQU毕业证书)成绩单原版一比一
如何办理麦考瑞大学毕业证(MQU毕业证书)成绩单原版一比一ozave
 
Is Your Volvo XC90 Displaying Anti-Skid Service Required Alert Here's Why
Is Your Volvo XC90 Displaying Anti-Skid Service Required Alert Here's WhyIs Your Volvo XC90 Displaying Anti-Skid Service Required Alert Here's Why
Is Your Volvo XC90 Displaying Anti-Skid Service Required Alert Here's WhyBavarium Autoworks
 
如何办理(NCL毕业证书)纽卡斯尔大学毕业证毕业证成绩单原版一比一
如何办理(NCL毕业证书)纽卡斯尔大学毕业证毕业证成绩单原版一比一如何办理(NCL毕业证书)纽卡斯尔大学毕业证毕业证成绩单原版一比一
如何办理(NCL毕业证书)纽卡斯尔大学毕业证毕业证成绩单原版一比一avy6anjnd
 
What Does The Engine Malfunction Reduced Power Message Mean For Your BMW X5
What Does The Engine Malfunction Reduced Power Message Mean For Your BMW X5What Does The Engine Malfunction Reduced Power Message Mean For Your BMW X5
What Does The Engine Malfunction Reduced Power Message Mean For Your BMW X5Bavarian Workshop
 
Connaught Place, Delhi Call girls :8448380779 Model Escorts | 100% verified
Connaught Place, Delhi Call girls :8448380779 Model Escorts | 100% verifiedConnaught Place, Delhi Call girls :8448380779 Model Escorts | 100% verified
Connaught Place, Delhi Call girls :8448380779 Model Escorts | 100% verifiedDelhi Call girls
 
Call Girls Kadugodi Just Call 👗 7737669865 👗 Top Class Call Girl Service Bang...
Call Girls Kadugodi Just Call 👗 7737669865 👗 Top Class Call Girl Service Bang...Call Girls Kadugodi Just Call 👗 7737669865 👗 Top Class Call Girl Service Bang...
Call Girls Kadugodi Just Call 👗 7737669865 👗 Top Class Call Girl Service Bang...amitlee9823
 
Call Girls In Kotla Mubarakpur Delhi ❤️8448577510 ⊹Best Escorts Service In 24...
Call Girls In Kotla Mubarakpur Delhi ❤️8448577510 ⊹Best Escorts Service In 24...Call Girls In Kotla Mubarakpur Delhi ❤️8448577510 ⊹Best Escorts Service In 24...
Call Girls In Kotla Mubarakpur Delhi ❤️8448577510 ⊹Best Escorts Service In 24...lizamodels9
 
Why Does My Porsche Cayenne's Exhaust Sound So Loud
Why Does My Porsche Cayenne's Exhaust Sound So LoudWhy Does My Porsche Cayenne's Exhaust Sound So Loud
Why Does My Porsche Cayenne's Exhaust Sound So LoudRoyalty Auto Service
 
Top Rated Call Girls Mumbai Central : 9920725232 We offer Beautiful and sexy ...
Top Rated Call Girls Mumbai Central : 9920725232 We offer Beautiful and sexy ...Top Rated Call Girls Mumbai Central : 9920725232 We offer Beautiful and sexy ...
Top Rated Call Girls Mumbai Central : 9920725232 We offer Beautiful and sexy ...amitlee9823
 
Madiwala Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...
Madiwala Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...Madiwala Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...
Madiwala Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...amitlee9823
 
Vip Mumbai Call Girls Mira Road Call On 9920725232 With Body to body massage ...
Vip Mumbai Call Girls Mira Road Call On 9920725232 With Body to body massage ...Vip Mumbai Call Girls Mira Road Call On 9920725232 With Body to body massage ...
Vip Mumbai Call Girls Mira Road Call On 9920725232 With Body to body massage ...amitlee9823
 
BOOK FARIDABAD CALL GIRL(VIP Sunny Leone) @8168257667 BOOK 24/7
BOOK FARIDABAD CALL GIRL(VIP Sunny Leone) @8168257667 BOOK  24/7BOOK FARIDABAD CALL GIRL(VIP Sunny Leone) @8168257667 BOOK  24/7
BOOK FARIDABAD CALL GIRL(VIP Sunny Leone) @8168257667 BOOK 24/7Hyderabad Escorts Agency
 
➥🔝 7737669865 🔝▻ Bhiwandi Call-girls in Women Seeking Men 🔝Bhiwandi🔝 Escor...
➥🔝 7737669865 🔝▻ Bhiwandi Call-girls in Women Seeking Men  🔝Bhiwandi🔝   Escor...➥🔝 7737669865 🔝▻ Bhiwandi Call-girls in Women Seeking Men  🔝Bhiwandi🔝   Escor...
➥🔝 7737669865 🔝▻ Bhiwandi Call-girls in Women Seeking Men 🔝Bhiwandi🔝 Escor...amitlee9823
 

Recently uploaded (20)

Somya Surve Escorts Service Bilaspur ❣️ 7014168258 ❣️ High Cost Unlimited Har...
Somya Surve Escorts Service Bilaspur ❣️ 7014168258 ❣️ High Cost Unlimited Har...Somya Surve Escorts Service Bilaspur ❣️ 7014168258 ❣️ High Cost Unlimited Har...
Somya Surve Escorts Service Bilaspur ❣️ 7014168258 ❣️ High Cost Unlimited Har...
 
FULL NIGHT — 9999894380 Call Girls In Jagat Puri | Delhi
FULL NIGHT — 9999894380 Call Girls In Jagat Puri | DelhiFULL NIGHT — 9999894380 Call Girls In Jagat Puri | Delhi
FULL NIGHT — 9999894380 Call Girls In Jagat Puri | Delhi
 
(INDIRA) Call Girl Nashik Call Now 8617697112 Nashik Escorts 24x7
(INDIRA) Call Girl Nashik Call Now 8617697112 Nashik Escorts 24x7(INDIRA) Call Girl Nashik Call Now 8617697112 Nashik Escorts 24x7
(INDIRA) Call Girl Nashik Call Now 8617697112 Nashik Escorts 24x7
 
Majestic Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...
Majestic Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...Majestic Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...
Majestic Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...
 
如何办理麦考瑞大学毕业证(MQU毕业证书)成绩单原版一比一
如何办理麦考瑞大学毕业证(MQU毕业证书)成绩单原版一比一如何办理麦考瑞大学毕业证(MQU毕业证书)成绩单原版一比一
如何办理麦考瑞大学毕业证(MQU毕业证书)成绩单原版一比一
 
(INDIRA) Call Girl Surat Call Now 8250077686 Surat Escorts 24x7
(INDIRA) Call Girl Surat Call Now 8250077686 Surat Escorts 24x7(INDIRA) Call Girl Surat Call Now 8250077686 Surat Escorts 24x7
(INDIRA) Call Girl Surat Call Now 8250077686 Surat Escorts 24x7
 
Is Your Volvo XC90 Displaying Anti-Skid Service Required Alert Here's Why
Is Your Volvo XC90 Displaying Anti-Skid Service Required Alert Here's WhyIs Your Volvo XC90 Displaying Anti-Skid Service Required Alert Here's Why
Is Your Volvo XC90 Displaying Anti-Skid Service Required Alert Here's Why
 
如何办理(NCL毕业证书)纽卡斯尔大学毕业证毕业证成绩单原版一比一
如何办理(NCL毕业证书)纽卡斯尔大学毕业证毕业证成绩单原版一比一如何办理(NCL毕业证书)纽卡斯尔大学毕业证毕业证成绩单原版一比一
如何办理(NCL毕业证书)纽卡斯尔大学毕业证毕业证成绩单原版一比一
 
What Does The Engine Malfunction Reduced Power Message Mean For Your BMW X5
What Does The Engine Malfunction Reduced Power Message Mean For Your BMW X5What Does The Engine Malfunction Reduced Power Message Mean For Your BMW X5
What Does The Engine Malfunction Reduced Power Message Mean For Your BMW X5
 
Connaught Place, Delhi Call girls :8448380779 Model Escorts | 100% verified
Connaught Place, Delhi Call girls :8448380779 Model Escorts | 100% verifiedConnaught Place, Delhi Call girls :8448380779 Model Escorts | 100% verified
Connaught Place, Delhi Call girls :8448380779 Model Escorts | 100% verified
 
Call Girls Kadugodi Just Call 👗 7737669865 👗 Top Class Call Girl Service Bang...
Call Girls Kadugodi Just Call 👗 7737669865 👗 Top Class Call Girl Service Bang...Call Girls Kadugodi Just Call 👗 7737669865 👗 Top Class Call Girl Service Bang...
Call Girls Kadugodi Just Call 👗 7737669865 👗 Top Class Call Girl Service Bang...
 
Call Girls In Kotla Mubarakpur Delhi ❤️8448577510 ⊹Best Escorts Service In 24...
Call Girls In Kotla Mubarakpur Delhi ❤️8448577510 ⊹Best Escorts Service In 24...Call Girls In Kotla Mubarakpur Delhi ❤️8448577510 ⊹Best Escorts Service In 24...
Call Girls In Kotla Mubarakpur Delhi ❤️8448577510 ⊹Best Escorts Service In 24...
 
Why Does My Porsche Cayenne's Exhaust Sound So Loud
Why Does My Porsche Cayenne's Exhaust Sound So LoudWhy Does My Porsche Cayenne's Exhaust Sound So Loud
Why Does My Porsche Cayenne's Exhaust Sound So Loud
 
Top Rated Call Girls Mumbai Central : 9920725232 We offer Beautiful and sexy ...
Top Rated Call Girls Mumbai Central : 9920725232 We offer Beautiful and sexy ...Top Rated Call Girls Mumbai Central : 9920725232 We offer Beautiful and sexy ...
Top Rated Call Girls Mumbai Central : 9920725232 We offer Beautiful and sexy ...
 
Madiwala Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...
Madiwala Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...Madiwala Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...
Madiwala Call Girls: 🍓 7737669865 🍓 High Profile Model Escorts | Bangalore Es...
 
Vip Mumbai Call Girls Mira Road Call On 9920725232 With Body to body massage ...
Vip Mumbai Call Girls Mira Road Call On 9920725232 With Body to body massage ...Vip Mumbai Call Girls Mira Road Call On 9920725232 With Body to body massage ...
Vip Mumbai Call Girls Mira Road Call On 9920725232 With Body to body massage ...
 
BOOK FARIDABAD CALL GIRL(VIP Sunny Leone) @8168257667 BOOK 24/7
BOOK FARIDABAD CALL GIRL(VIP Sunny Leone) @8168257667 BOOK  24/7BOOK FARIDABAD CALL GIRL(VIP Sunny Leone) @8168257667 BOOK  24/7
BOOK FARIDABAD CALL GIRL(VIP Sunny Leone) @8168257667 BOOK 24/7
 
(ISHITA) Call Girls Service Jammu Call Now 8617697112 Jammu Escorts 24x7
(ISHITA) Call Girls Service Jammu Call Now 8617697112 Jammu Escorts 24x7(ISHITA) Call Girls Service Jammu Call Now 8617697112 Jammu Escorts 24x7
(ISHITA) Call Girls Service Jammu Call Now 8617697112 Jammu Escorts 24x7
 
Stay Cool and Compliant: Know Your Window Tint Laws Before You Tint
Stay Cool and Compliant: Know Your Window Tint Laws Before You TintStay Cool and Compliant: Know Your Window Tint Laws Before You Tint
Stay Cool and Compliant: Know Your Window Tint Laws Before You Tint
 
➥🔝 7737669865 🔝▻ Bhiwandi Call-girls in Women Seeking Men 🔝Bhiwandi🔝 Escor...
➥🔝 7737669865 🔝▻ Bhiwandi Call-girls in Women Seeking Men  🔝Bhiwandi🔝   Escor...➥🔝 7737669865 🔝▻ Bhiwandi Call-girls in Women Seeking Men  🔝Bhiwandi🔝   Escor...
➥🔝 7737669865 🔝▻ Bhiwandi Call-girls in Women Seeking Men 🔝Bhiwandi🔝 Escor...
 

Deep learning state_of_the_art- Autonomous Driving

  • 1. Deep Learning: State of the Art (2019)
  • 2. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: State of the Art* (Breakthrough Developments in 2017 & 2018) • BERT and Natural Language Processing • Tesla Autopilot Hardware v2+: Neural Networks at Scale • AdaNet: AutoML with Ensembles • AutoAugment: Deep RL Data Augmentation • Training Deep Networks with Synthetic Data • Segmentation Annotation with Polygon-RNN++ • DAWNBench: Training Fast and Cheap • BigGAN: State of the Art in Image Synthesis • Video-to-Video Synthesis • Semantic Segmentation • AlphaZero & OpenAI Five • Deep Learning Frameworks * This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an overview of exciting recent developments.
  • 3. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Encoder-Decoder Architecture Sequence-to-Sequence Model - Neural Machine Translation [295] Encoder RNN encodes input sequence into a fixed size vector, and then is passed repeatedly to decoder RNN.
  • 4. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Attention [296] Attention mechanism allows the network to refer back to the input sequence, instead of forcing it to encode all information into one fixed-length vector.
  • 5. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Attention [312]
  • 6. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Attention [312]
  • 7. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Self-Attention [312] "The animal didn't cross the street because it was too tired" More details: http://jalammar.github.io/illustrated-transformer/
  • 8. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Transformer [312] Vaswani, Ashish, et al. "Attention is all you need." Advances in Neural Information Processing Systems. 2017.
  • 9. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references Word Embeddings (Word2Vec) [297] Word Vector Skip Gram Model:
  • 10. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Context-Aware Embeddings [312]
  • 11. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references OpenAI Transformer 1. Pre-train a Transformer’s decoder for language modeling 2. Train it on, for example, a sentence classification task [312] 1 2
  • 12. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references BERT Devlin, Jacob, et al. "Bert: Pre-training of deep bidirectional transformers for language understanding." (2018).
  • 13. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references BERT Applications Now you can use BERT: • Create contextualized word embeddings (like ELMo) • Sentence classification • Sentence pair classification • Sentence pair similarity • Multiple choice • Sentence tagging • Question answering
  • 14. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: State of the Art* (Breakthrough Developments in 2017 & 2018) • BERT and Natural Language Processing • Tesla Autopilot Hardware v2+: Neural Networks at Scale • AdaNet: AutoML with Ensembles • AutoAugment: Deep RL Data Augmentation • Training Deep Networks with Synthetic Data • Segmentation Annotation with Polygon-RNN++ • DAWNBench: Training Fast and Cheap • BigGAN: State of the Art in Image Synthesis • Video-to-Video Synthesis • Semantic Segmentation • AlphaZero & OpenAI Five • Deep Learning Frameworks * This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an overview of exciting recent developments.
  • 15. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Tesla Autopilot Hardware v2+ • Specialized NVIDIA Drive PX 2 hardware • Neural network takes all 8 cameras as input • Based on Inception v1 architecture
  • 16. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Autopilot Reaches 1 Billion Miles (~0.5 Billion on Hardware v2+)
  • 17. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: State of the Art* (Breakthrough Developments in 2017 & 2018) • BERT and Natural Language Processing • Tesla Autopilot Hardware v2+: Neural Networks at Scale • AdaNet: AutoML with Ensembles • AutoAugment: Deep RL Data Augmentation • Training Deep Networks with Synthetic Data • Segmentation Annotation with Polygon-RNN++ • DAWNBench: Training Fast and Cheap • BigGAN: State of the Art in Image Synthesis • Video-to-Video Synthesis • Semantic Segmentation • AlphaZero & OpenAI Five • Deep Learning Frameworks * This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an overview of exciting recent developments.
  • 18. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references AutoML and Neural Architecture Search (NASNet) [300, 301]
  • 19. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references AdaNet [317]
  • 20. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: State of the Art* (Breakthrough Developments in 2017 & 2018) • BERT and Natural Language Processing • Tesla Autopilot Hardware v2+: Neural Networks at Scale • AdaNet: AutoML with Ensembles • AutoAugment: Deep RL Data Augmentation • Training Deep Networks with Synthetic Data • Segmentation Annotation with Polygon-RNN++ • DAWNBench: Training Fast and Cheap • BigGAN: State of the Art in Image Synthesis • Video-to-Video Synthesis • Semantic Segmentation • AlphaZero & OpenAI Five • Deep Learning Frameworks * This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an overview of exciting recent developments.
  • 21. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references AutoAugment: RL for Data Augmentation [318] • Show that transfer learning can also be done augmentation policies instead of weights (or with).
  • 22. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: State of the Art* (Breakthrough Developments in 2017 & 2018) • BERT and Natural Language Processing • Tesla Autopilot Hardware v2+: Neural Networks at Scale • AdaNet: AutoML with Ensembles • AutoAugment: Deep RL Data Augmentation • Training Deep Networks with Synthetic Data • Segmentation Annotation with Polygon-RNN++ • DAWNBench: Training Fast and Cheap • BigGAN: State of the Art in Image Synthesis • Video-to-Video Synthesis • Semantic Segmentation • AlphaZero & OpenAI Five • Deep Learning Frameworks * This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an overview of exciting recent developments.
  • 23. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Training on Randomized Synthetic Data [319] • number and types of objects • number, types, colors, and scales of distractors • texture on the object of interest, and background photograph • location of the virtual camera with respect to the scene • angle of the camera with respect to the scene • number and locations of point lights Tremblay, Jonathan, et al. "Training deep networks with synthetic data: Bridging the reality gap by domain randomization." (2018).
  • 24. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: State of the Art* (Breakthrough Developments in 2017 & 2018) • BERT and Natural Language Processing • Tesla Autopilot Hardware v2+: Neural Networks at Scale • AdaNet: AutoML with Ensembles • AutoAugment: Deep RL Data Augmentation • Training Deep Networks with Synthetic Data • Segmentation Annotation with Polygon-RNN++ • DAWNBench: Training Fast and Cheap • BigGAN: State of the Art in Image Synthesis • Video-to-Video Synthesis • Semantic Segmentation • AlphaZero & OpenAI Five • Deep Learning Frameworks * This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an overview of exciting recent developments.
  • 25. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Segmentation Annotation with Polygon-RNN++ • High-resolution polygon segmentation generation with human-in-the-loop [319] Acuna, David, et al. "Efficient Interactive Annotation of Segmentation Datasets With Polygon-RNN++." CVPR 2018.
  • 26. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: State of the Art* (Breakthrough Developments in 2017 & 2018) • BERT and Natural Language Processing • Tesla Autopilot Hardware v2+: Neural Networks at Scale • AdaNet: AutoML with Ensembles • AutoAugment: Deep RL Data Augmentation • Training Deep Networks with Synthetic Data • Segmentation Annotation with Polygon-RNN++ • DAWNBench: Training Fast and Cheap • BigGAN: State of the Art in Image Synthesis • Video-to-Video Synthesis • Semantic Segmentation • AlphaZero & OpenAI Five • Deep Learning Frameworks * This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an overview of exciting recent developments.
  • 27. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references DAWNBench: Training Fast and Cheap DAWNBench: • Competition on speed and cost of training and inference that achieves 93% for ImageNet and 94% for CIFAR 10 fast.ai – Training: • ImageNet in 3 hours for $25 • CIFAR10 for $0.26 • Key idea: During training, if you very slowly increase learning rate while decreasing momentum, you can train at extremely high learning rates, thus avoiding over-fitting, and training in far fewer epochs. • Details: http://bit.ly/2H6yv6H [320]
  • 28. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: State of the Art* (Breakthrough Developments in 2017 & 2018) • BERT and Natural Language Processing • Tesla Autopilot Hardware v2+: Neural Networks at Scale • AdaNet: AutoML with Ensembles • AutoAugment: Deep RL Data Augmentation • Training Deep Networks with Synthetic Data • Segmentation Annotation with Polygon-RNN++ • DAWNBench: Training Fast and Cheap • BigGAN: State of the Art in Image Synthesis • Video-to-Video Synthesis • Semantic Segmentation • AlphaZero & OpenAI Five • Deep Learning Frameworks * This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an overview of exciting recent developments.
  • 29. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references BigGAN: State of the Art in Image Synthesis • Same GAN techniques, much larger scale • Increase model capacity + increase batch size [321] Brock, Andrew, Jeff Donahue, and Karen Simonyan. "Large scale gan training for high fidelity natural image synthesis." (2018).
  • 30. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: State of the Art* (Breakthrough Developments in 2017 & 2018) • BERT and Natural Language Processing • Tesla Autopilot Hardware v2+: Neural Networks at Scale • AdaNet: AutoML with Ensembles • AutoAugment: Deep RL Data Augmentation • Training Deep Networks with Synthetic Data • Segmentation Annotation with Polygon-RNN++ • DAWNBench: Training Fast and Cheap • BigGAN: State of the Art in Image Synthesis • Video-to-Video Synthesis • Semantic Segmentation • AlphaZero & OpenAI Five • Deep Learning Frameworks * This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an overview of exciting recent developments.
  • 31. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Video-to-Video Synthesis • Model temporal dynamics by adding a spatio-temporal adversarial objective to the GAN framework [322] Wang, Ting-Chun, et al. "Video-to-video synthesis." (2018).
  • 32. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Video-to-Video Synthesis • Model temporal dynamics by adding a spatio-temporal adversarial objective to the GAN framework [322] Wang, Ting-Chun, et al. "Video-to-video synthesis." (2018).
  • 33. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Video-to-Video Synthesis • Model temporal dynamics by adding a spatio-temporal adversarial objective to the GAN framework [322] Wang, Ting-Chun, et al. "Video-to-video synthesis." (2018).
  • 34. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: State of the Art* (Breakthrough Developments in 2017 & 2018) • BERT and Natural Language Processing • Tesla Autopilot Hardware v2+: Neural Networks at Scale • AdaNet: AutoML with Ensembles • AutoAugment: Deep RL Data Augmentation • Training Deep Networks with Synthetic Data • Segmentation Annotation with Polygon-RNN++ • DAWNBench: Training Fast and Cheap • BigGAN: State of the Art in Image Synthesis • Video-to-Video Synthesis • Semantic Segmentation • AlphaZero & OpenAI Five • Deep Learning Frameworks * This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an overview of exciting recent developments.
  • 35. 2019https://deeplearning.mit.edu • AlexNet (2012): First CNN (15.4%) • 8 layers • 61 million parameters • ZFNet (2013): 15.4% to 11.2% • 8 layers • More filters. Denser stride. • VGGNet (2014): 11.2% to 7.3% • Beautifully uniform: 3x3 conv, stride 1, pad 1, 2x2 max pool • 16 layers • 138 million parameters • GoogLeNet (2014): 11.2% to 6.7% • Inception modules • 22 layers • 5 million parameters (throw away fully connected layers) • ResNet (2015): 6.7% to 3.57% • More layers = better performance • 152 layers • CUImage (2016): 3.57% to 2.99% • Ensemble of 6 models • SENet (2017): 2.99% to 2.251% • Squeeze and excitation block: network is allowed to adaptively adjust the weighting of each feature map in the convolutional block. References: [90] Human error (5.1%) surpassed in 2015
  • 36. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Object Detection / Localization Region-Based Methods | Shown: Faster R-CNN [299]
  • 37. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Object Detection / Localization Single-Shot Methods | Shown: SSD [299]
  • 38. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Object Detection: State of the Art Progress [326]
  • 39. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Semantic Segmentation [175]
  • 40. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references State-of-the-Art: DeepLab v3 • PASCAL VOC Challenge: http://bit.ly/2HdzTEu
  • 41. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references FCN (Nov 2014) Paper: “Fully Convolutional Networks for Semantic Segmentation” • Repurpose Imagenet pretrained nets • Upsample using deconvolution • Skip connections to improve coarseness of upsampling [173]
  • 42. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references SegNet (Nov 2015) Paper: “SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation” • Maxpooling indices transferred to decoder to improve the segmentation resolution. [175]
  • 43. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references Dilated Convolutions (Nov 2015) Paper: “Multi-Scale Context Aggregation by Dilated Convolutions” • Since pooling decreases resolution: • Added “dilated convolution layer” • Still interpolate up from 1/8 of original image size [175]
  • 44. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references DeepLab v1, v2 (Jun 2016) Paper: “DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs” • Added fully-connected Conditional Random Fields (CRFs) – as a post-processing step • Smooth segmentation based on the underlying image intensities [179]
  • 45. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references Key Aspects of Segmentation • Fully convolutional networks (FCNs) - replace fully-connected layers with convolutional layers. Classification network is where the biggest gains come from. • Deeper, updated models (ResNet, etc) consistent with ImageNet Challenge object classification tasks. • Conditional Random Fields (CRFs) to capture both local and long-range dependencies within an image to refine the prediction map. • Dilated convolution (aka Atrous convolution) – maintain computational cost, increase resolution of intermediate feature maps • Process at multiple scales and combine the information together [174, 176]
  • 46. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references DeepLab v3 Paper: “Rethinking Atrous Convolution for Semantic Image Segmentation” • Multi-scale processing, without increasing parameters. • Increasing “atrous rate” enlarges the model’s “field-of-view” [179]
  • 47. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references DeepLab v3 trained on CityScapes Tutorial: https://github.com/lexfridman/mit-deep-learning
  • 48. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: State of the Art* (Breakthrough Developments in 2017 & 2018) • BERT and Natural Language Processing • Tesla Autopilot Hardware v2+: Neural Networks at Scale • AdaNet: AutoML with Ensembles • AutoAugment: Deep RL Data Augmentation • Training Deep Networks with Synthetic Data • Segmentation Annotation with Polygon-RNN++ • DAWNBench: Training Fast and Cheap • BigGAN: State of the Art in Image Synthesis • Video-to-Video Synthesis • Semantic Segmentation • AlphaZero & OpenAI Five • Deep Learning Frameworks * This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an overview of exciting recent developments.
  • 49. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references Atari Breakout [85] After 120 Minutes of Training After 10 Minutes of Training After 240 Minutes of Training
  • 50. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references DQN: Atari Games (2015) [83]
  • 51. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references AlphaGo (2016): Beat Top Human at Go [83]
  • 52. 2019https://deeplearning.mit.eduFor the full updated list of references visit: https://selfdrivingcars.mit.edu/references AlphaGo Zero (2017): Beats AlphaGo [149]
  • 53. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references [324] AlphaZero (Dec 2017) vs StockFish (Chess) & Elmo (Shogi)
  • 54. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references OpenAI & Dota 2 • Dota 2 as a testbed for the messiness and continuous nature of the real world: teamwork, long time horizons, and hidden information. [325]
  • 55. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references OpenAI & Dota 2 Progress • Aug, 2017: 1v1 bot beats top professional Dota 2 players. • Aug, 2018: OpenAI Five lost two games against top Dota 2 players at The International. “We are looking forward to pushing Five to the next level.” [325]
  • 56. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: State of the Art* (Breakthrough Developments in 2017 & 2018) • BERT and Natural Language Processing • Tesla Autopilot Hardware v2+: Neural Networks at Scale • AdaNet: AutoML with Ensembles • AutoAugment: Deep RL Data Augmentation • Training Deep Networks with Synthetic Data • Segmentation Annotation with Polygon-RNN++ • DAWNBench: Training Fast and Cheap • BigGAN: State of the Art in Image Synthesis • Video-to-Video Synthesis • Semantic Segmentation • AlphaZero & OpenAI Five • Deep Learning Frameworks * This is not a list of state-of-the-art results on main machine learning benchmark datasets. It’s an overview of exciting recent developments.
  • 57. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning Frameworks [327] Deep Learning Framework Power Scores (by Jeff Hale) http://bit.ly/2GBa3tU 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. Factors to consider: • Learning curve • Speed of development • Size and passion of community • Number of papers implemented in framework • Likelihood of long-term growth and stability • Ecosystem of tooling
  • 58. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Deep Learning: 2019 and Beyond • On backpropagation: “My view is throw it all away and start again.” • “The future depends on some graduate student who is deeply suspicious of everything I have said.” - Geoffrey Hinton “Godfather of Deep Learning” [328]
  • 59. 2019https://deeplearning.mit.eduFor the full list of references visit: https://hcai.mit.edu/references Thank You Website: deeplearning.mit.edu • Videos and slides posted on the website • Code posted on GitHub: https://github.com/lexfridman/mit-deep-learning