SlideShare a Scribd company logo
Continual Learning over Small non-i.i.d. Batches
of Natural Video Streams
ContinualAI May Meetup
Gabriele Graffieti
gabriele.graffieti@unibo.it
PhD Student
Computer Science and Engineering Department
University of Bologna
Italy
May 29, 2020
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 1 / 18
Sequential data
A CL ideal application
• A new object is shown to a CL agent.
• The agent acquire a short video of the object.
• Frame extracted from the video constitute one or more small mini-batches, containing a
highly correlated patterns from single class.
Problems
• standard SGD-based optimization does not work well on this setting.
• High forgetting, since a single batch contains only one object.
• In popular benchmarks (CIFAR, ImageNet) instances of the same class are independent.
Its unlikely that an application experiences at time t a set of independent images of the same
class. It’s more realistic to encounter a single object and observe it under different poses.
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 2 / 18
Our setting
CORe50
• 50 classes, 10 categories 5 different objects per category.
• ∼165,000 128×128 RGB-D images.
• for each object there are 11 different video sessions (∼300 frames recorded with a Kinect
2 at 20 fps).
NICv2-391
• New Instances and Classes scenario.
• 391 batches of 300 images each (first batch 10 times larger).
• Only one object per incremental batch.
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 3 / 18
NICv2-391
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 4 / 18
AR1*
Convolutionallayers
Head
Data layer
For each batch:
• weights in the online head are zero-init for new
classes and reload from off-line head for known
classes.
• at the end of the batch weights are consolidated in
the off-line memory by weighted averaging.
cw[j] =
cw[j] · wpastj + (tw[j] − avg(tw))
wpastj + 1
wpastj =
pastj
curj
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 5 / 18
AR1*
Convolutionallayers
Head
Data layer
• To control forgetting in the lower layers we use
Synaptic Intelligence (limit update of important
weights).
• SI exploits information made available by SGD,
does not require any further gradient propagation.
• Weight update does not require to store the old
weights.
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 6 / 18
Results on CORe50 NICv2
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 7 / 18
Results on CORe50 NICv2
Strat. Run time (m) Data (MB) Params. (MB)
CWR* 21.4 0 0.2
Naive 25.6 0 0
LWF 27.8 0 0
EWC 31.2 0 24.4
AR1* 39.9 0 12.2
DSLDA 79.1 0 0.2
Cumul. 2826.2 5,898.3 0
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 8 / 18
Improve the results with replay
Convolutionallayers
Head
Data layer Replay memory
• Straightforward technique: just store past
data and repeat them through the
network.
• No need to weight constraints in the
convolutional layers (AR1*free).
• If few images per class are taken, not so
critical impact on memory overhead.
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 9 / 18
Improve the results with replay
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 10 / 18
Replay problems
Convolutionallayers
Head
Data layer Replay memory
• Requires extra storage (e.g. for ImageNet,
if we store 20 patterns per class, the total
storage is about 3.8 GB)
• Requires extra forward/backward steps
when mixing new and old patterns more
iterations for epoch.
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 11 / 18
Latent replay
Advantages:
• Efficiency: extra forward and backward
steps take place only in the upper layers.
• Less storage required.
• Activations can be quantized/compressed
with negligible accuracy loss.
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 12 / 18
Latent replay
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 13 / 18
Strategy Run time (m) Mem. overhead (MB) Final Accuracy
Naive 25.6 0 + 0 7.13%
CWR* 21.4 0 + 0.2 56.99%
DSLDA 79.1 0 + 0.2 48.02%
AR1* 39.9 0 + 12.4 56.32%
AR1*free (Image) 133.3 75 + 0 77.30%
AR1*free (conv5 4/dw) 41.2 48 + 0 72.23%
AR1*free (pool6) 23.7 5.8 + 0 59.75%
Cumulative 2826.2 ∼6,000 + 0 85.26%
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 14 / 18
Future works
• Latent Generative Replay.
• Self-training by exploiting temporal coherence.
• Openset classification (automatic discovery of new classes).
• Sparse human supervision (active learning).
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 15 / 18
Thank You!
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 16 / 18
Bibliography I
Vincenzo Lomonaco, Davide Maltoni
CORe50: a new Dataset and Benchmark for continual Object Recognition
CoRL 2017
Vincenzo Lomonaco, Davide Maltoni, Lorenzo Pellegrini
Rehearsal-Free Continual Learning over Small Non-I.I.D. Batches
CVPR 2020 (Workshop)
Li Zhizhong, Derek Hoiem
Learning without Forgetting
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 17 / 18
Bibliography II
James Kirkpatrick et al.
Overcoming catastrophic forgetting in neural networks
Proceedings of the National Academy of Sciences, 2017
Tyler L. Hayes, Christopher Kanan
Lifelong Machine Learning with Deep Streaming Linear Discriminant Analysis
arXiv Preprint, 2019
Lorenzo Pellegrini, Gabriele Graffieti, Vincenzo Lomonaco, Davide Maltoni
Latent Replay for Real-Time Continual Learning
arXiv Preprint, 2020
Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 18 / 18

More Related Content

What's hot

Deep Learning Representations for All (a.ka. the AI hype)
Deep Learning Representations for All (a.ka. the AI hype)Deep Learning Representations for All (a.ka. the AI hype)
Deep Learning Representations for All (a.ka. the AI hype)
Universitat Politècnica de Catalunya
 
Multimodal Deep Learning
Multimodal Deep LearningMultimodal Deep Learning
Multimodal Deep Learning
Universitat Politècnica de Catalunya
 
Deep Learning: a birds eye view
Deep Learning: a birds eye viewDeep Learning: a birds eye view
Deep Learning: a birds eye view
Roelof Pieters
 
Deep Learning: concepts and use cases (October 2018)
Deep Learning: concepts and use cases (October 2018)Deep Learning: concepts and use cases (October 2018)
Deep Learning: concepts and use cases (October 2018)
Julien SIMON
 
Transfer Learning: Breve introducción a modelos pre-entrenados.
Transfer Learning: Breve introducción a modelos pre-entrenados.Transfer Learning: Breve introducción a modelos pre-entrenados.
Transfer Learning: Breve introducción a modelos pre-entrenados.
Fernando Constantino
 
Neural networks and deep learning
Neural networks and deep learningNeural networks and deep learning
Neural networks and deep learning
Jörgen Sandig
 
2a Mini-conf PredictCovid. Field: Artificial Intelligence
2a Mini-conf PredictCovid. Field: Artificial Intelligence2a Mini-conf PredictCovid. Field: Artificial Intelligence
2a Mini-conf PredictCovid. Field: Artificial Intelligence
Alex Camargo
 
Deep Video Object Tracking 2020 - Xavier Giro - UPC TelecomBCN Barcelona
Deep Video Object Tracking 2020 - Xavier Giro - UPC TelecomBCN BarcelonaDeep Video Object Tracking 2020 - Xavier Giro - UPC TelecomBCN Barcelona
Deep Video Object Tracking 2020 - Xavier Giro - UPC TelecomBCN Barcelona
Universitat Politècnica de Catalunya
 
Recurrent Neural Networks RNN - Xavier Giro - UPC TelecomBCN Barcelona 2020
Recurrent Neural Networks RNN - Xavier Giro - UPC TelecomBCN Barcelona 2020Recurrent Neural Networks RNN - Xavier Giro - UPC TelecomBCN Barcelona 2020
Recurrent Neural Networks RNN - Xavier Giro - UPC TelecomBCN Barcelona 2020
Universitat Politècnica de Catalunya
 
Yann le cun
Yann le cunYann le cun
Yann le cunYandex
 
Deep Learning for Artificial Intelligence (AI)
Deep Learning for Artificial Intelligence (AI)Deep Learning for Artificial Intelligence (AI)
Deep Learning for Artificial Intelligence (AI)
Er. Shiva K. Shrestha
 
What's Wrong With Deep Learning?
What's Wrong With Deep Learning?What's Wrong With Deep Learning?
What's Wrong With Deep Learning?
Philip Zheng
 
How machines can take decisions
How machines can take decisionsHow machines can take decisions
How machines can take decisions
Deepu S Nath
 
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
Siby Jose Plathottam
 
Advance deep learning
Advance deep learningAdvance deep learning
Advance deep learning
aliaKhan71
 
Multimodal Learning Analytics
Multimodal Learning AnalyticsMultimodal Learning Analytics
Multimodal Learning Analytics
Xavier Ochoa
 

What's hot (17)

Deep Learning Representations for All (a.ka. the AI hype)
Deep Learning Representations for All (a.ka. the AI hype)Deep Learning Representations for All (a.ka. the AI hype)
Deep Learning Representations for All (a.ka. the AI hype)
 
Multimodal Deep Learning
Multimodal Deep LearningMultimodal Deep Learning
Multimodal Deep Learning
 
Deep Learning: a birds eye view
Deep Learning: a birds eye viewDeep Learning: a birds eye view
Deep Learning: a birds eye view
 
Deep Learning: concepts and use cases (October 2018)
Deep Learning: concepts and use cases (October 2018)Deep Learning: concepts and use cases (October 2018)
Deep Learning: concepts and use cases (October 2018)
 
DEFENSE
DEFENSEDEFENSE
DEFENSE
 
Transfer Learning: Breve introducción a modelos pre-entrenados.
Transfer Learning: Breve introducción a modelos pre-entrenados.Transfer Learning: Breve introducción a modelos pre-entrenados.
Transfer Learning: Breve introducción a modelos pre-entrenados.
 
Neural networks and deep learning
Neural networks and deep learningNeural networks and deep learning
Neural networks and deep learning
 
2a Mini-conf PredictCovid. Field: Artificial Intelligence
2a Mini-conf PredictCovid. Field: Artificial Intelligence2a Mini-conf PredictCovid. Field: Artificial Intelligence
2a Mini-conf PredictCovid. Field: Artificial Intelligence
 
Deep Video Object Tracking 2020 - Xavier Giro - UPC TelecomBCN Barcelona
Deep Video Object Tracking 2020 - Xavier Giro - UPC TelecomBCN BarcelonaDeep Video Object Tracking 2020 - Xavier Giro - UPC TelecomBCN Barcelona
Deep Video Object Tracking 2020 - Xavier Giro - UPC TelecomBCN Barcelona
 
Recurrent Neural Networks RNN - Xavier Giro - UPC TelecomBCN Barcelona 2020
Recurrent Neural Networks RNN - Xavier Giro - UPC TelecomBCN Barcelona 2020Recurrent Neural Networks RNN - Xavier Giro - UPC TelecomBCN Barcelona 2020
Recurrent Neural Networks RNN - Xavier Giro - UPC TelecomBCN Barcelona 2020
 
Yann le cun
Yann le cunYann le cun
Yann le cun
 
Deep Learning for Artificial Intelligence (AI)
Deep Learning for Artificial Intelligence (AI)Deep Learning for Artificial Intelligence (AI)
Deep Learning for Artificial Intelligence (AI)
 
What's Wrong With Deep Learning?
What's Wrong With Deep Learning?What's Wrong With Deep Learning?
What's Wrong With Deep Learning?
 
How machines can take decisions
How machines can take decisionsHow machines can take decisions
How machines can take decisions
 
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
 
Advance deep learning
Advance deep learningAdvance deep learning
Advance deep learning
 
Multimodal Learning Analytics
Multimodal Learning AnalyticsMultimodal Learning Analytics
Multimodal Learning Analytics
 

Similar to Continual Learning over Small non-i.i.d. Batches of Natural Video Streams

Video summarization using clustering
Video summarization using clusteringVideo summarization using clustering
Video summarization using clustering
Sahil Biswas
 
(Paper Review)U-GAT-IT: unsupervised generative attentional networks with ada...
(Paper Review)U-GAT-IT: unsupervised generative attentional networks with ada...(Paper Review)U-GAT-IT: unsupervised generative attentional networks with ada...
(Paper Review)U-GAT-IT: unsupervised generative attentional networks with ada...
MYEONGGYU LEE
 
PR-433: Test-time Training with Masked Autoencoders
PR-433: Test-time Training with Masked AutoencodersPR-433: Test-time Training with Masked Autoencoders
PR-433: Test-time Training with Masked Autoencoders
Sunghoon Joo
 
IRJET- Storage Optimization of Video Surveillance from CCTV Camera
IRJET- Storage Optimization of Video Surveillance from CCTV CameraIRJET- Storage Optimization of Video Surveillance from CCTV Camera
IRJET- Storage Optimization of Video Surveillance from CCTV Camera
IRJET Journal
 
Multimedia Buzzword Bingo: Translating to English
 Multimedia Buzzword Bingo: Translating to English Multimedia Buzzword Bingo: Translating to English
Multimedia Buzzword Bingo: Translating to English
Derek Buitenhuis
 
BigML Education - Deepnets
BigML Education - DeepnetsBigML Education - Deepnets
BigML Education - Deepnets
BigML, Inc
 
Paper discussion:Video-to-Video Synthesis (NIPS 2018)
Paper discussion:Video-to-Video Synthesis (NIPS 2018)Paper discussion:Video-to-Video Synthesis (NIPS 2018)
Paper discussion:Video-to-Video Synthesis (NIPS 2018)
Motaz Sabri
 
SOFI Developer Meeting Göttingen 28th March 2015
SOFI Developer MeetingGöttingen 28th March 2015 SOFI Developer MeetingGöttingen 28th March 2015
SOFI Developer Meeting Göttingen 28th March 2015
Dirk Hähnel
 
“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...
“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...
“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...
Edge AI and Vision Alliance
 
Masked Feature Modelling for the unsupervised pre-training of a Graph Attenti...
Masked Feature Modelling for the unsupervised pre-training of a Graph Attenti...Masked Feature Modelling for the unsupervised pre-training of a Graph Attenti...
Masked Feature Modelling for the unsupervised pre-training of a Graph Attenti...
VasileiosMezaris
 
Gated-ViGAT
Gated-ViGATGated-ViGAT
Gated-ViGAT
VasileiosMezaris
 
AI Food detector; A model of Generative adversarial network for food Classifier
AI Food detector; A model of Generative adversarial network for food ClassifierAI Food detector; A model of Generative adversarial network for food Classifier
AI Food detector; A model of Generative adversarial network for food Classifier
jimmy majumder
 
CNN Quantization
CNN QuantizationCNN Quantization
CNN Quantization
Emanuele Ghelfi
 
PERFORMANCEATMOSPHERE - PREDICTION OF GPU-BASED DEEP LEARNING APPLICATIONS
PERFORMANCEATMOSPHERE - PREDICTION OF GPU-BASED DEEP LEARNING APPLICATIONSPERFORMANCEATMOSPHERE - PREDICTION OF GPU-BASED DEEP LEARNING APPLICATIONS
PERFORMANCEATMOSPHERE - PREDICTION OF GPU-BASED DEEP LEARNING APPLICATIONS
ATMOSPHERE .
 
IRJET- Survey Paper on Anomaly Detection in Surveillance Videos
IRJET-  	  Survey Paper on Anomaly Detection in Surveillance VideosIRJET-  	  Survey Paper on Anomaly Detection in Surveillance Videos
IRJET- Survey Paper on Anomaly Detection in Surveillance Videos
IRJET Journal
 
GNorm and Rethinking pre training-ruijie
GNorm and Rethinking pre training-ruijieGNorm and Rethinking pre training-ruijie
GNorm and Rethinking pre training-ruijie
哲东 郑
 
Csw2016 d antoine_automatic_exploitgeneration
Csw2016 d antoine_automatic_exploitgenerationCsw2016 d antoine_automatic_exploitgeneration
Csw2016 d antoine_automatic_exploitgeneration
CanSecWest
 
2019-06-14:3 - Reti neurali e compressione video
2019-06-14:3 - Reti neurali e compressione video2019-06-14:3 - Reti neurali e compressione video
2019-06-14:3 - Reti neurali e compressione video
uninfoit
 
IRJET- A Hybrid Image and Video Compression of DCT and DWT Techniques for H.2...
IRJET- A Hybrid Image and Video Compression of DCT and DWT Techniques for H.2...IRJET- A Hybrid Image and Video Compression of DCT and DWT Techniques for H.2...
IRJET- A Hybrid Image and Video Compression of DCT and DWT Techniques for H.2...
IRJET Journal
 
Explaining video summarization based on the focus of attention
Explaining video summarization based on the focus of attentionExplaining video summarization based on the focus of attention
Explaining video summarization based on the focus of attention
VasileiosMezaris
 

Similar to Continual Learning over Small non-i.i.d. Batches of Natural Video Streams (20)

Video summarization using clustering
Video summarization using clusteringVideo summarization using clustering
Video summarization using clustering
 
(Paper Review)U-GAT-IT: unsupervised generative attentional networks with ada...
(Paper Review)U-GAT-IT: unsupervised generative attentional networks with ada...(Paper Review)U-GAT-IT: unsupervised generative attentional networks with ada...
(Paper Review)U-GAT-IT: unsupervised generative attentional networks with ada...
 
PR-433: Test-time Training with Masked Autoencoders
PR-433: Test-time Training with Masked AutoencodersPR-433: Test-time Training with Masked Autoencoders
PR-433: Test-time Training with Masked Autoencoders
 
IRJET- Storage Optimization of Video Surveillance from CCTV Camera
IRJET- Storage Optimization of Video Surveillance from CCTV CameraIRJET- Storage Optimization of Video Surveillance from CCTV Camera
IRJET- Storage Optimization of Video Surveillance from CCTV Camera
 
Multimedia Buzzword Bingo: Translating to English
 Multimedia Buzzword Bingo: Translating to English Multimedia Buzzword Bingo: Translating to English
Multimedia Buzzword Bingo: Translating to English
 
BigML Education - Deepnets
BigML Education - DeepnetsBigML Education - Deepnets
BigML Education - Deepnets
 
Paper discussion:Video-to-Video Synthesis (NIPS 2018)
Paper discussion:Video-to-Video Synthesis (NIPS 2018)Paper discussion:Video-to-Video Synthesis (NIPS 2018)
Paper discussion:Video-to-Video Synthesis (NIPS 2018)
 
SOFI Developer Meeting Göttingen 28th March 2015
SOFI Developer MeetingGöttingen 28th March 2015 SOFI Developer MeetingGöttingen 28th March 2015
SOFI Developer Meeting Göttingen 28th March 2015
 
“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...
“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...
“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...
 
Masked Feature Modelling for the unsupervised pre-training of a Graph Attenti...
Masked Feature Modelling for the unsupervised pre-training of a Graph Attenti...Masked Feature Modelling for the unsupervised pre-training of a Graph Attenti...
Masked Feature Modelling for the unsupervised pre-training of a Graph Attenti...
 
Gated-ViGAT
Gated-ViGATGated-ViGAT
Gated-ViGAT
 
AI Food detector; A model of Generative adversarial network for food Classifier
AI Food detector; A model of Generative adversarial network for food ClassifierAI Food detector; A model of Generative adversarial network for food Classifier
AI Food detector; A model of Generative adversarial network for food Classifier
 
CNN Quantization
CNN QuantizationCNN Quantization
CNN Quantization
 
PERFORMANCEATMOSPHERE - PREDICTION OF GPU-BASED DEEP LEARNING APPLICATIONS
PERFORMANCEATMOSPHERE - PREDICTION OF GPU-BASED DEEP LEARNING APPLICATIONSPERFORMANCEATMOSPHERE - PREDICTION OF GPU-BASED DEEP LEARNING APPLICATIONS
PERFORMANCEATMOSPHERE - PREDICTION OF GPU-BASED DEEP LEARNING APPLICATIONS
 
IRJET- Survey Paper on Anomaly Detection in Surveillance Videos
IRJET-  	  Survey Paper on Anomaly Detection in Surveillance VideosIRJET-  	  Survey Paper on Anomaly Detection in Surveillance Videos
IRJET- Survey Paper on Anomaly Detection in Surveillance Videos
 
GNorm and Rethinking pre training-ruijie
GNorm and Rethinking pre training-ruijieGNorm and Rethinking pre training-ruijie
GNorm and Rethinking pre training-ruijie
 
Csw2016 d antoine_automatic_exploitgeneration
Csw2016 d antoine_automatic_exploitgenerationCsw2016 d antoine_automatic_exploitgeneration
Csw2016 d antoine_automatic_exploitgeneration
 
2019-06-14:3 - Reti neurali e compressione video
2019-06-14:3 - Reti neurali e compressione video2019-06-14:3 - Reti neurali e compressione video
2019-06-14:3 - Reti neurali e compressione video
 
IRJET- A Hybrid Image and Video Compression of DCT and DWT Techniques for H.2...
IRJET- A Hybrid Image and Video Compression of DCT and DWT Techniques for H.2...IRJET- A Hybrid Image and Video Compression of DCT and DWT Techniques for H.2...
IRJET- A Hybrid Image and Video Compression of DCT and DWT Techniques for H.2...
 
Explaining video summarization based on the focus of attention
Explaining video summarization based on the focus of attentionExplaining video summarization based on the focus of attention
Explaining video summarization based on the focus of attention
 

More from Gabriele Graffieti

Introduction to AI Ethics
Introduction to AI EthicsIntroduction to AI Ethics
Introduction to AI Ethics
Gabriele Graffieti
 
Goals, Risks and Countermeasures in the Artificial Intelligence Era
Goals, Risks and Countermeasures in the Artificial Intelligence EraGoals, Risks and Countermeasures in the Artificial Intelligence Era
Goals, Risks and Countermeasures in the Artificial Intelligence Era
Gabriele Graffieti
 
Efficient Continual Learning with Latent Rehearsal
Efficient Continual Learning with Latent RehearsalEfficient Continual Learning with Latent Rehearsal
Efficient Continual Learning with Latent Rehearsal
Gabriele Graffieti
 
Spotlight - Generative Models and GANs
Spotlight - Generative Models and GANsSpotlight - Generative Models and GANs
Spotlight - Generative Models and GANs
Gabriele Graffieti
 
From art to deepfakes: an introduction to Generative Adversarial Networks
From art to deepfakes: an introduction to Generative Adversarial NetworksFrom art to deepfakes: an introduction to Generative Adversarial Networks
From art to deepfakes: an introduction to Generative Adversarial Networks
Gabriele Graffieti
 
Image-to-image Translation with Generative Adversarial Networks (without math)
Image-to-image Translation with Generative Adversarial Networks (without math)Image-to-image Translation with Generative Adversarial Networks (without math)
Image-to-image Translation with Generative Adversarial Networks (without math)
Gabriele Graffieti
 

More from Gabriele Graffieti (6)

Introduction to AI Ethics
Introduction to AI EthicsIntroduction to AI Ethics
Introduction to AI Ethics
 
Goals, Risks and Countermeasures in the Artificial Intelligence Era
Goals, Risks and Countermeasures in the Artificial Intelligence EraGoals, Risks and Countermeasures in the Artificial Intelligence Era
Goals, Risks and Countermeasures in the Artificial Intelligence Era
 
Efficient Continual Learning with Latent Rehearsal
Efficient Continual Learning with Latent RehearsalEfficient Continual Learning with Latent Rehearsal
Efficient Continual Learning with Latent Rehearsal
 
Spotlight - Generative Models and GANs
Spotlight - Generative Models and GANsSpotlight - Generative Models and GANs
Spotlight - Generative Models and GANs
 
From art to deepfakes: an introduction to Generative Adversarial Networks
From art to deepfakes: an introduction to Generative Adversarial NetworksFrom art to deepfakes: an introduction to Generative Adversarial Networks
From art to deepfakes: an introduction to Generative Adversarial Networks
 
Image-to-image Translation with Generative Adversarial Networks (without math)
Image-to-image Translation with Generative Adversarial Networks (without math)Image-to-image Translation with Generative Adversarial Networks (without math)
Image-to-image Translation with Generative Adversarial Networks (without math)
 

Recently uploaded

Elevating Tactical DDD Patterns Through Object Calisthenics
Elevating Tactical DDD Patterns Through Object CalisthenicsElevating Tactical DDD Patterns Through Object Calisthenics
Elevating Tactical DDD Patterns Through Object Calisthenics
Dorra BARTAGUIZ
 
FIDO Alliance Osaka Seminar: FIDO Security Aspects.pdf
FIDO Alliance Osaka Seminar: FIDO Security Aspects.pdfFIDO Alliance Osaka Seminar: FIDO Security Aspects.pdf
FIDO Alliance Osaka Seminar: FIDO Security Aspects.pdf
FIDO Alliance
 
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
DanBrown980551
 
State of ICS and IoT Cyber Threat Landscape Report 2024 preview
State of ICS and IoT Cyber Threat Landscape Report 2024 previewState of ICS and IoT Cyber Threat Landscape Report 2024 preview
State of ICS and IoT Cyber Threat Landscape Report 2024 preview
Prayukth K V
 
From Siloed Products to Connected Ecosystem: Building a Sustainable and Scala...
From Siloed Products to Connected Ecosystem: Building a Sustainable and Scala...From Siloed Products to Connected Ecosystem: Building a Sustainable and Scala...
From Siloed Products to Connected Ecosystem: Building a Sustainable and Scala...
Product School
 
The Future of Platform Engineering
The Future of Platform EngineeringThe Future of Platform Engineering
The Future of Platform Engineering
Jemma Hussein Allen
 
GraphRAG is All You need? LLM & Knowledge Graph
GraphRAG is All You need? LLM & Knowledge GraphGraphRAG is All You need? LLM & Knowledge Graph
GraphRAG is All You need? LLM & Knowledge Graph
Guy Korland
 
Accelerate your Kubernetes clusters with Varnish Caching
Accelerate your Kubernetes clusters with Varnish CachingAccelerate your Kubernetes clusters with Varnish Caching
Accelerate your Kubernetes clusters with Varnish Caching
Thijs Feryn
 
Bits & Pixels using AI for Good.........
Bits & Pixels using AI for Good.........Bits & Pixels using AI for Good.........
Bits & Pixels using AI for Good.........
Alison B. Lowndes
 
Transcript: Selling digital books in 2024: Insights from industry leaders - T...
Transcript: Selling digital books in 2024: Insights from industry leaders - T...Transcript: Selling digital books in 2024: Insights from industry leaders - T...
Transcript: Selling digital books in 2024: Insights from industry leaders - T...
BookNet Canada
 
Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024
Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024
Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024
Tobias Schneck
 
Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...
Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...
Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...
Thierry Lestable
 
Smart TV Buyer Insights Survey 2024 by 91mobiles.pdf
Smart TV Buyer Insights Survey 2024 by 91mobiles.pdfSmart TV Buyer Insights Survey 2024 by 91mobiles.pdf
Smart TV Buyer Insights Survey 2024 by 91mobiles.pdf
91mobiles
 
JMeter webinar - integration with InfluxDB and Grafana
JMeter webinar - integration with InfluxDB and GrafanaJMeter webinar - integration with InfluxDB and Grafana
JMeter webinar - integration with InfluxDB and Grafana
RTTS
 
FIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdf
FIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdfFIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdf
FIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdf
FIDO Alliance
 
Builder.ai Founder Sachin Dev Duggal's Strategic Approach to Create an Innova...
Builder.ai Founder Sachin Dev Duggal's Strategic Approach to Create an Innova...Builder.ai Founder Sachin Dev Duggal's Strategic Approach to Create an Innova...
Builder.ai Founder Sachin Dev Duggal's Strategic Approach to Create an Innova...
Ramesh Iyer
 
FIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdf
FIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdfFIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdf
FIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdf
FIDO Alliance
 
Monitoring Java Application Security with JDK Tools and JFR Events
Monitoring Java Application Security with JDK Tools and JFR EventsMonitoring Java Application Security with JDK Tools and JFR Events
Monitoring Java Application Security with JDK Tools and JFR Events
Ana-Maria Mihalceanu
 
Slack (or Teams) Automation for Bonterra Impact Management (fka Social Soluti...
Slack (or Teams) Automation for Bonterra Impact Management (fka Social Soluti...Slack (or Teams) Automation for Bonterra Impact Management (fka Social Soluti...
Slack (or Teams) Automation for Bonterra Impact Management (fka Social Soluti...
Jeffrey Haguewood
 
Essentials of Automations: Optimizing FME Workflows with Parameters
Essentials of Automations: Optimizing FME Workflows with ParametersEssentials of Automations: Optimizing FME Workflows with Parameters
Essentials of Automations: Optimizing FME Workflows with Parameters
Safe Software
 

Recently uploaded (20)

Elevating Tactical DDD Patterns Through Object Calisthenics
Elevating Tactical DDD Patterns Through Object CalisthenicsElevating Tactical DDD Patterns Through Object Calisthenics
Elevating Tactical DDD Patterns Through Object Calisthenics
 
FIDO Alliance Osaka Seminar: FIDO Security Aspects.pdf
FIDO Alliance Osaka Seminar: FIDO Security Aspects.pdfFIDO Alliance Osaka Seminar: FIDO Security Aspects.pdf
FIDO Alliance Osaka Seminar: FIDO Security Aspects.pdf
 
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
 
State of ICS and IoT Cyber Threat Landscape Report 2024 preview
State of ICS and IoT Cyber Threat Landscape Report 2024 previewState of ICS and IoT Cyber Threat Landscape Report 2024 preview
State of ICS and IoT Cyber Threat Landscape Report 2024 preview
 
From Siloed Products to Connected Ecosystem: Building a Sustainable and Scala...
From Siloed Products to Connected Ecosystem: Building a Sustainable and Scala...From Siloed Products to Connected Ecosystem: Building a Sustainable and Scala...
From Siloed Products to Connected Ecosystem: Building a Sustainable and Scala...
 
The Future of Platform Engineering
The Future of Platform EngineeringThe Future of Platform Engineering
The Future of Platform Engineering
 
GraphRAG is All You need? LLM & Knowledge Graph
GraphRAG is All You need? LLM & Knowledge GraphGraphRAG is All You need? LLM & Knowledge Graph
GraphRAG is All You need? LLM & Knowledge Graph
 
Accelerate your Kubernetes clusters with Varnish Caching
Accelerate your Kubernetes clusters with Varnish CachingAccelerate your Kubernetes clusters with Varnish Caching
Accelerate your Kubernetes clusters with Varnish Caching
 
Bits & Pixels using AI for Good.........
Bits & Pixels using AI for Good.........Bits & Pixels using AI for Good.........
Bits & Pixels using AI for Good.........
 
Transcript: Selling digital books in 2024: Insights from industry leaders - T...
Transcript: Selling digital books in 2024: Insights from industry leaders - T...Transcript: Selling digital books in 2024: Insights from industry leaders - T...
Transcript: Selling digital books in 2024: Insights from industry leaders - T...
 
Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024
Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024
Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024
 
Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...
Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...
Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...
 
Smart TV Buyer Insights Survey 2024 by 91mobiles.pdf
Smart TV Buyer Insights Survey 2024 by 91mobiles.pdfSmart TV Buyer Insights Survey 2024 by 91mobiles.pdf
Smart TV Buyer Insights Survey 2024 by 91mobiles.pdf
 
JMeter webinar - integration with InfluxDB and Grafana
JMeter webinar - integration with InfluxDB and GrafanaJMeter webinar - integration with InfluxDB and Grafana
JMeter webinar - integration with InfluxDB and Grafana
 
FIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdf
FIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdfFIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdf
FIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdf
 
Builder.ai Founder Sachin Dev Duggal's Strategic Approach to Create an Innova...
Builder.ai Founder Sachin Dev Duggal's Strategic Approach to Create an Innova...Builder.ai Founder Sachin Dev Duggal's Strategic Approach to Create an Innova...
Builder.ai Founder Sachin Dev Duggal's Strategic Approach to Create an Innova...
 
FIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdf
FIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdfFIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdf
FIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdf
 
Monitoring Java Application Security with JDK Tools and JFR Events
Monitoring Java Application Security with JDK Tools and JFR EventsMonitoring Java Application Security with JDK Tools and JFR Events
Monitoring Java Application Security with JDK Tools and JFR Events
 
Slack (or Teams) Automation for Bonterra Impact Management (fka Social Soluti...
Slack (or Teams) Automation for Bonterra Impact Management (fka Social Soluti...Slack (or Teams) Automation for Bonterra Impact Management (fka Social Soluti...
Slack (or Teams) Automation for Bonterra Impact Management (fka Social Soluti...
 
Essentials of Automations: Optimizing FME Workflows with Parameters
Essentials of Automations: Optimizing FME Workflows with ParametersEssentials of Automations: Optimizing FME Workflows with Parameters
Essentials of Automations: Optimizing FME Workflows with Parameters
 

Continual Learning over Small non-i.i.d. Batches of Natural Video Streams

  • 1. Continual Learning over Small non-i.i.d. Batches of Natural Video Streams ContinualAI May Meetup Gabriele Graffieti gabriele.graffieti@unibo.it PhD Student Computer Science and Engineering Department University of Bologna Italy May 29, 2020 Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 1 / 18
  • 2. Sequential data A CL ideal application • A new object is shown to a CL agent. • The agent acquire a short video of the object. • Frame extracted from the video constitute one or more small mini-batches, containing a highly correlated patterns from single class. Problems • standard SGD-based optimization does not work well on this setting. • High forgetting, since a single batch contains only one object. • In popular benchmarks (CIFAR, ImageNet) instances of the same class are independent. Its unlikely that an application experiences at time t a set of independent images of the same class. It’s more realistic to encounter a single object and observe it under different poses. Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 2 / 18
  • 3. Our setting CORe50 • 50 classes, 10 categories 5 different objects per category. • ∼165,000 128×128 RGB-D images. • for each object there are 11 different video sessions (∼300 frames recorded with a Kinect 2 at 20 fps). NICv2-391 • New Instances and Classes scenario. • 391 batches of 300 images each (first batch 10 times larger). • Only one object per incremental batch. Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 3 / 18
  • 4. NICv2-391 Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 4 / 18
  • 5. AR1* Convolutionallayers Head Data layer For each batch: • weights in the online head are zero-init for new classes and reload from off-line head for known classes. • at the end of the batch weights are consolidated in the off-line memory by weighted averaging. cw[j] = cw[j] · wpastj + (tw[j] − avg(tw)) wpastj + 1 wpastj = pastj curj Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 5 / 18
  • 6. AR1* Convolutionallayers Head Data layer • To control forgetting in the lower layers we use Synaptic Intelligence (limit update of important weights). • SI exploits information made available by SGD, does not require any further gradient propagation. • Weight update does not require to store the old weights. Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 6 / 18
  • 7. Results on CORe50 NICv2 Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 7 / 18
  • 8. Results on CORe50 NICv2 Strat. Run time (m) Data (MB) Params. (MB) CWR* 21.4 0 0.2 Naive 25.6 0 0 LWF 27.8 0 0 EWC 31.2 0 24.4 AR1* 39.9 0 12.2 DSLDA 79.1 0 0.2 Cumul. 2826.2 5,898.3 0 Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 8 / 18
  • 9. Improve the results with replay Convolutionallayers Head Data layer Replay memory • Straightforward technique: just store past data and repeat them through the network. • No need to weight constraints in the convolutional layers (AR1*free). • If few images per class are taken, not so critical impact on memory overhead. Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 9 / 18
  • 10. Improve the results with replay Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 10 / 18
  • 11. Replay problems Convolutionallayers Head Data layer Replay memory • Requires extra storage (e.g. for ImageNet, if we store 20 patterns per class, the total storage is about 3.8 GB) • Requires extra forward/backward steps when mixing new and old patterns more iterations for epoch. Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 11 / 18
  • 12. Latent replay Advantages: • Efficiency: extra forward and backward steps take place only in the upper layers. • Less storage required. • Activations can be quantized/compressed with negligible accuracy loss. Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 12 / 18
  • 13. Latent replay Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 13 / 18
  • 14. Strategy Run time (m) Mem. overhead (MB) Final Accuracy Naive 25.6 0 + 0 7.13% CWR* 21.4 0 + 0.2 56.99% DSLDA 79.1 0 + 0.2 48.02% AR1* 39.9 0 + 12.4 56.32% AR1*free (Image) 133.3 75 + 0 77.30% AR1*free (conv5 4/dw) 41.2 48 + 0 72.23% AR1*free (pool6) 23.7 5.8 + 0 59.75% Cumulative 2826.2 ∼6,000 + 0 85.26% Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 14 / 18
  • 15. Future works • Latent Generative Replay. • Self-training by exploiting temporal coherence. • Openset classification (automatic discovery of new classes). • Sparse human supervision (active learning). Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 15 / 18
  • 16. Thank You! Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 16 / 18
  • 17. Bibliography I Vincenzo Lomonaco, Davide Maltoni CORe50: a new Dataset and Benchmark for continual Object Recognition CoRL 2017 Vincenzo Lomonaco, Davide Maltoni, Lorenzo Pellegrini Rehearsal-Free Continual Learning over Small Non-I.I.D. Batches CVPR 2020 (Workshop) Li Zhizhong, Derek Hoiem Learning without Forgetting IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017 Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 17 / 18
  • 18. Bibliography II James Kirkpatrick et al. Overcoming catastrophic forgetting in neural networks Proceedings of the National Academy of Sciences, 2017 Tyler L. Hayes, Christopher Kanan Lifelong Machine Learning with Deep Streaming Linear Discriminant Analysis arXiv Preprint, 2019 Lorenzo Pellegrini, Gabriele Graffieti, Vincenzo Lomonaco, Davide Maltoni Latent Replay for Real-Time Continual Learning arXiv Preprint, 2020 Gabriele Graffieti CL over Small non-i.i.d. Batches of Video Streams May 29, 2020 18 / 18