SlideShare a Scribd company logo
,
Don’t forget, there is more than forgetting: new metrics for Continual Learning
Natalia D´ıaz-Rodr´ıguez1∗
, Vincenzo Lomonaco2∗
, Davide Maltoni2
and David Filliat1
,
(1) U2IS, ENSTA ParisTech, Inria FLOWERS team, Palaiseau, France. https://flowers.inria.fr/ {natalia.diaz, david.filliat}@ensta-paristech.fr.
(2) University of Bologna, Italy. {vincenzo.lomonaco, davide.maltoni}@unibo.it http://continualai.org/
(*) Both authors contributed equally to this work.
,
Motivation
The lack of consensus in evaluating continual learning (CL) algorithms and
the almost exclusive focus on catastrophic forgetting motivate us to propose a
more comprehensive set of implementation independent metrics accounting for
factors we believe have practical implications worth considering with respect
to “static” machine learning settings.
New metrics for Continual Learning (CL)
A CL algorithm ACL
is an algorithm with the following signature:
∀Di ∈ D, ACL
i : < hi−1, Tri, Mi−1, t >→< hi, Mi > (1)
Where hi is the model, Tri is the training set of examples drawn from the
respective Di distribution, Mi is an external memory where we can store
previous training examples and t is a task label. For simplicity, we can assume
N as the number of tasks, one for each Tri.
Accuracy (A): Given matrix R ∈ RN×N
, which contains in each entry Ri,j
the test classification accuracy of the model on task tj after observing the
last sample from task ti [3], Accuracy metric is:
A =
N
i≥j Ri,j
N(N+1)
2
(2)
Table 1: Accuracy matrix R: elements accounted to compute A (white &
cyan), BWT (cyan), and FWT (gray) Tri = training, Tei= test tasks.
R Te1 Te2 Te3
Tr1 R∗
Rij Rij
Tr2 Rij R∗
Rij
Tr3 Rij Rij R∗
Backward Transfer (BWT) measures the influence that learning a task has
on the performance on previous tasks [3]:
BWT =
N
i=2
i−1
j=1(Ri,j − Rj,j)
N(N−1)
2
(3)
Remembering (REM):
REM = 1 − |min(BWT, 0)| (4)
is the originally negative BWT while the originally positive BWT, i.e.,
improvement over time, is:
Positive Backward Transfer (BWT+
):
BWT+
= max(BWT, 0) (5)
Forward Transfer (FWT): measures the influence that learning a task
has on the performance of future tasks [3].
FWT =
N
i<j Ri,j
N(N−1)
2
(6)
Model size efficiency (MS): The memory size of model hi quantified in
terms of parameters θ at each task i, Mem(θi), should not grow too rapidly
with respect to the size of the model that learned the first task, Mem(θ1).
MS = min(1,
N
i=1
Mem(θ1)
Mem(θi)
N
) (7)
Samples storage size efficiency (SSS): The memory occupation in bits
by the samples storage memory M, Mem(M), should be bounded by the
memory occupation of the total nr of examples encountered at the end of
the last task (D is the lifetime dataset associated to all distributions D):
SSS = 1 − min(1,
N
i=1
Mem(Mi)
Mem(D)
N
) (8)
Computational efficiency (CE): it is bounded by the nr of operations
for training set Tri. Ops(Tri) is the number of (mul-adds) operations needed
to learn Tri and Ops ↑↓(Tri) is the nr of operations required to do one
forward and one backward (backprop) pass on Tri).
CE = min(1,
N
i=1
Ops↑↓(Tri)·ε
1+Ops(Tri)
N
) (9)
CLscore and CLstability aggregating metrics
CLscore: if ci ∈ [0, 1]) is the avg. of r runs assigned a weight wi ∈ [0, 1]
s.t. C
i wi = 1:
CLscore =
#C
i=1
wici (10)
CLstability: the average of the std. devs. from all previous criteria ci:
CLstability = 1 −
#C
i=1
wistddev(ci) (11)
Experiments and Conclusion
Figure 1: a) Spider chart: CL metrics per strategy (larger area is better). b)
Accuracy per CL strategy computed over the fixed test set as proposed in [4].
We evaluate the CL metrics on cumulative and na¨ıve baseline strategies as in
[4], Elastic Weight Consolidation (EWC) [1], Synaptic Intelligence (SI) [5] and
Learning without Forgetting (LwF) [2] on iCIFAR-100 dataset.
Table 2: CL metrics and CLscore for each CL strategy evaluated (higher is better).
Str. A REM BWT+
FWT MS SSS CE CLsco CLstab
Na¨ı 0.3825 0.6664 0.00 0.10 1.00 1.00 0.4492 0.5140 0.9986
Cum 0.7225 1.00 0.0673 0.10 1.00 0.55 0.1496 0.5128 0.9979
EWC 0.5940 0.9821 0.00 0.10 0.40 1.00 0.3495 0.4894 0.9972
LWF 0.5278 0.9667 0.00 0.10 1.00 1.00 0.4429 0.5768 0.9986
SI 0.5795 0.9620 0.00 0.10 0.40 1.00 0.3613 0.4861 0.9970
• Experiments: 3 weight configs
W =[wA, wMS, wSSS, wCE, wREM, wBWT+, wFWT ] (here wi = 1
#C, CNN in
[5, 4]).
• Future work: provide insights that assess the importance of different
metric schemes and their entanglement, and how to use these metrics wisely
to assist choosing among algorithms.
References
[1] J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A.
Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, D. Hassabis,
C. Clopath, D. Kumaran, and R. Hadsell.
Overcoming catastrophic forgetting in neural networks.
Proceedings of the National Academy of Sciences, 114(13):3521–3526, 2017.
[2] Z. Li and D. Hoiem.
Learning without forgetting.
IEEE Transactions on Pattern Analysis and Machine Intelligence,
40(12):2935–2947, Dec 2018.
[3] D. Lopez-Paz et al.
Gradient episodic memory for continual learning.
In Advances in Neural Information Processing Systems, pages 6467–6476,
2017.
[4] D. Maltoni and V. Lomonaco.
Continuous learning in single-incremental-task scenarios.
arXiv preprint arXiv:1806.08568, 2018.
[5] F. Zenke, B. Poole, and S. Ganguli.
Continual learning through synaptic intelligence.
In Proceedings of the 34th International Conference on Machine Learning,
Proceedings of Machine Learning Research, pages 3987–3995. PMLR, 2017.

More Related Content

What's hot

Data Science in #mHealth
Data Science in #mHealthData Science in #mHealth
Data Science in #mHealth
Neal Lathia
 
A STRATEGIC HYBRID TECHNIQUE TO DEVELOP A GAME PLAYER
A STRATEGIC HYBRID TECHNIQUE TO DEVELOP A GAME PLAYERA STRATEGIC HYBRID TECHNIQUE TO DEVELOP A GAME PLAYER
A STRATEGIC HYBRID TECHNIQUE TO DEVELOP A GAME PLAYER
ijcseit
 
Machine Learning: Make Your Ruby Code Smarter
Machine Learning: Make Your Ruby Code SmarterMachine Learning: Make Your Ruby Code Smarter
Machine Learning: Make Your Ruby Code Smarter
Astrails
 
Designed by Identity MLP
Designed by Identity MLP Designed by Identity MLP
Designed by Identity MLP butest
 
Fixed-Point Code Synthesis for Neural Networks
Fixed-Point Code Synthesis for Neural NetworksFixed-Point Code Synthesis for Neural Networks
Fixed-Point Code Synthesis for Neural Networks
gerogepatton
 
International Journal of Computational Engineering Research(IJCER)
International Journal of Computational Engineering Research(IJCER)International Journal of Computational Engineering Research(IJCER)
International Journal of Computational Engineering Research(IJCER)ijceronline
 

What's hot (6)

Data Science in #mHealth
Data Science in #mHealthData Science in #mHealth
Data Science in #mHealth
 
A STRATEGIC HYBRID TECHNIQUE TO DEVELOP A GAME PLAYER
A STRATEGIC HYBRID TECHNIQUE TO DEVELOP A GAME PLAYERA STRATEGIC HYBRID TECHNIQUE TO DEVELOP A GAME PLAYER
A STRATEGIC HYBRID TECHNIQUE TO DEVELOP A GAME PLAYER
 
Machine Learning: Make Your Ruby Code Smarter
Machine Learning: Make Your Ruby Code SmarterMachine Learning: Make Your Ruby Code Smarter
Machine Learning: Make Your Ruby Code Smarter
 
Designed by Identity MLP
Designed by Identity MLP Designed by Identity MLP
Designed by Identity MLP
 
Fixed-Point Code Synthesis for Neural Networks
Fixed-Point Code Synthesis for Neural NetworksFixed-Point Code Synthesis for Neural Networks
Fixed-Point Code Synthesis for Neural Networks
 
International Journal of Computational Engineering Research(IJCER)
International Journal of Computational Engineering Research(IJCER)International Journal of Computational Engineering Research(IJCER)
International Journal of Computational Engineering Research(IJCER)
 

Similar to Don't forget, there is more than forgetting: new metrics for Continual Learning - Poster

Integration of Unsupervised and Supervised Criteria for DNNs Training
Integration of Unsupervised and Supervised Criteria for DNNs TrainingIntegration of Unsupervised and Supervised Criteria for DNNs Training
Integration of Unsupervised and Supervised Criteria for DNNs Training
Francisco Zamora-Martinez
 
Tim connecting-the-dots
Tim connecting-the-dotsTim connecting-the-dots
Tim connecting-the-dots
Timothy Head
 
Anirban part1
Anirban part1Anirban part1
Anirban part1
kamatchi priya
 
Data-Driven Recommender Systems
Data-Driven Recommender SystemsData-Driven Recommender Systems
Data-Driven Recommender Systems
recsysfr
 
Chap 8. Optimization for training deep models
Chap 8. Optimization for training deep modelsChap 8. Optimization for training deep models
Chap 8. Optimization for training deep models
Young-Geun Choi
 
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic...
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic...Quantization and Training of Neural Networks for Efficient Integer-Arithmetic...
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic...
Ryo Takahashi
 
TOWARDS MORE ACCURATE CLUSTERING METHOD BY USING DYNAMIC TIME WARPING
TOWARDS MORE ACCURATE CLUSTERING METHOD BY USING DYNAMIC TIME WARPINGTOWARDS MORE ACCURATE CLUSTERING METHOD BY USING DYNAMIC TIME WARPING
TOWARDS MORE ACCURATE CLUSTERING METHOD BY USING DYNAMIC TIME WARPING
ijdkp
 
[DOLAP2023] The Whys and Wherefores of Cubes
[DOLAP2023] The Whys and Wherefores of Cubes[DOLAP2023] The Whys and Wherefores of Cubes
[DOLAP2023] The Whys and Wherefores of Cubes
University of Bologna
 
Faster Interleaved Modular Multiplier Based on Sign Detection
Faster Interleaved Modular Multiplier Based on Sign DetectionFaster Interleaved Modular Multiplier Based on Sign Detection
Faster Interleaved Modular Multiplier Based on Sign Detection
VLSICS Design
 
Methodological Study Of Opinion Mining And Sentiment Analysis Techniques
Methodological Study Of Opinion Mining And Sentiment Analysis Techniques  Methodological Study Of Opinion Mining And Sentiment Analysis Techniques
Methodological Study Of Opinion Mining And Sentiment Analysis Techniques
ijsc
 
IndabaX Ghana Poster.pdf
IndabaX Ghana Poster.pdfIndabaX Ghana Poster.pdf
IndabaX Ghana Poster.pdf
kwadwoAmedi
 
Big Data & Machine Learning - TDC2013 São Paulo - 12/0713
Big Data & Machine Learning - TDC2013 São Paulo - 12/0713Big Data & Machine Learning - TDC2013 São Paulo - 12/0713
Big Data & Machine Learning - TDC2013 São Paulo - 12/0713
Mathieu DESPRIEE
 
Methodological study of opinion mining and sentiment analysis techniques
Methodological study of opinion mining and sentiment analysis techniquesMethodological study of opinion mining and sentiment analysis techniques
Methodological study of opinion mining and sentiment analysis techniques
ijsc
 
Big Data & Machine Learning - TDC2013 Sao Paulo
Big Data & Machine Learning - TDC2013 Sao PauloBig Data & Machine Learning - TDC2013 Sao Paulo
Big Data & Machine Learning - TDC2013 Sao Paulo
OCTO Technology
 
Lecture 1.pptx
Lecture 1.pptxLecture 1.pptx
Visualization of sorting algorithms using flash
Visualization of sorting algorithms using flashVisualization of sorting algorithms using flash
Visualization of sorting algorithms using flashiaemedu
 
Visualization of sorting algorithms using flash
Visualization of sorting algorithms using flashVisualization of sorting algorithms using flash
Visualization of sorting algorithms using flashiaemedu
 
Accelerated Particle Swarm Optimization and Support Vector Machine for Busine...
Accelerated Particle Swarm Optimization and Support Vector Machine for Busine...Accelerated Particle Swarm Optimization and Support Vector Machine for Busine...
Accelerated Particle Swarm Optimization and Support Vector Machine for Busine...
Xin-She Yang
 
Documentaries use for the design of learning activities
Documentaries use for the design of learning activitiesDocumentaries use for the design of learning activities
Documentaries use for the design of learning activities
IOSR Journals
 
IRJET - A Survey on Machine Learning Algorithms, Techniques and Applications
IRJET - A Survey on Machine Learning Algorithms, Techniques and ApplicationsIRJET - A Survey on Machine Learning Algorithms, Techniques and Applications
IRJET - A Survey on Machine Learning Algorithms, Techniques and Applications
IRJET Journal
 

Similar to Don't forget, there is more than forgetting: new metrics for Continual Learning - Poster (20)

Integration of Unsupervised and Supervised Criteria for DNNs Training
Integration of Unsupervised and Supervised Criteria for DNNs TrainingIntegration of Unsupervised and Supervised Criteria for DNNs Training
Integration of Unsupervised and Supervised Criteria for DNNs Training
 
Tim connecting-the-dots
Tim connecting-the-dotsTim connecting-the-dots
Tim connecting-the-dots
 
Anirban part1
Anirban part1Anirban part1
Anirban part1
 
Data-Driven Recommender Systems
Data-Driven Recommender SystemsData-Driven Recommender Systems
Data-Driven Recommender Systems
 
Chap 8. Optimization for training deep models
Chap 8. Optimization for training deep modelsChap 8. Optimization for training deep models
Chap 8. Optimization for training deep models
 
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic...
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic...Quantization and Training of Neural Networks for Efficient Integer-Arithmetic...
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic...
 
TOWARDS MORE ACCURATE CLUSTERING METHOD BY USING DYNAMIC TIME WARPING
TOWARDS MORE ACCURATE CLUSTERING METHOD BY USING DYNAMIC TIME WARPINGTOWARDS MORE ACCURATE CLUSTERING METHOD BY USING DYNAMIC TIME WARPING
TOWARDS MORE ACCURATE CLUSTERING METHOD BY USING DYNAMIC TIME WARPING
 
[DOLAP2023] The Whys and Wherefores of Cubes
[DOLAP2023] The Whys and Wherefores of Cubes[DOLAP2023] The Whys and Wherefores of Cubes
[DOLAP2023] The Whys and Wherefores of Cubes
 
Faster Interleaved Modular Multiplier Based on Sign Detection
Faster Interleaved Modular Multiplier Based on Sign DetectionFaster Interleaved Modular Multiplier Based on Sign Detection
Faster Interleaved Modular Multiplier Based on Sign Detection
 
Methodological Study Of Opinion Mining And Sentiment Analysis Techniques
Methodological Study Of Opinion Mining And Sentiment Analysis Techniques  Methodological Study Of Opinion Mining And Sentiment Analysis Techniques
Methodological Study Of Opinion Mining And Sentiment Analysis Techniques
 
IndabaX Ghana Poster.pdf
IndabaX Ghana Poster.pdfIndabaX Ghana Poster.pdf
IndabaX Ghana Poster.pdf
 
Big Data & Machine Learning - TDC2013 São Paulo - 12/0713
Big Data & Machine Learning - TDC2013 São Paulo - 12/0713Big Data & Machine Learning - TDC2013 São Paulo - 12/0713
Big Data & Machine Learning - TDC2013 São Paulo - 12/0713
 
Methodological study of opinion mining and sentiment analysis techniques
Methodological study of opinion mining and sentiment analysis techniquesMethodological study of opinion mining and sentiment analysis techniques
Methodological study of opinion mining and sentiment analysis techniques
 
Big Data & Machine Learning - TDC2013 Sao Paulo
Big Data & Machine Learning - TDC2013 Sao PauloBig Data & Machine Learning - TDC2013 Sao Paulo
Big Data & Machine Learning - TDC2013 Sao Paulo
 
Lecture 1.pptx
Lecture 1.pptxLecture 1.pptx
Lecture 1.pptx
 
Visualization of sorting algorithms using flash
Visualization of sorting algorithms using flashVisualization of sorting algorithms using flash
Visualization of sorting algorithms using flash
 
Visualization of sorting algorithms using flash
Visualization of sorting algorithms using flashVisualization of sorting algorithms using flash
Visualization of sorting algorithms using flash
 
Accelerated Particle Swarm Optimization and Support Vector Machine for Busine...
Accelerated Particle Swarm Optimization and Support Vector Machine for Busine...Accelerated Particle Swarm Optimization and Support Vector Machine for Busine...
Accelerated Particle Swarm Optimization and Support Vector Machine for Busine...
 
Documentaries use for the design of learning activities
Documentaries use for the design of learning activitiesDocumentaries use for the design of learning activities
Documentaries use for the design of learning activities
 
IRJET - A Survey on Machine Learning Algorithms, Techniques and Applications
IRJET - A Survey on Machine Learning Algorithms, Techniques and ApplicationsIRJET - A Survey on Machine Learning Algorithms, Techniques and Applications
IRJET - A Survey on Machine Learning Algorithms, Techniques and Applications
 

More from Vincenzo Lomonaco

2023-08-22 CoLLAs Tutorial - Beyond CIL.pdf
2023-08-22 CoLLAs Tutorial - Beyond CIL.pdf2023-08-22 CoLLAs Tutorial - Beyond CIL.pdf
2023-08-22 CoLLAs Tutorial - Beyond CIL.pdf
Vincenzo Lomonaco
 
Continual Learning with Deep Architectures - Tutorial ICML 2021
Continual Learning with Deep Architectures - Tutorial ICML 2021Continual Learning with Deep Architectures - Tutorial ICML 2021
Continual Learning with Deep Architectures - Tutorial ICML 2021
Vincenzo Lomonaco
 
Toward Continual Learning on the Edge
Toward Continual Learning on the EdgeToward Continual Learning on the Edge
Toward Continual Learning on the Edge
Vincenzo Lomonaco
 
Continual Learning: Another Step Towards Truly Intelligent Machines
Continual Learning: Another Step Towards Truly Intelligent MachinesContinual Learning: Another Step Towards Truly Intelligent Machines
Continual Learning: Another Step Towards Truly Intelligent Machines
Vincenzo Lomonaco
 
Tutorial inns2019 full
Tutorial inns2019 fullTutorial inns2019 full
Tutorial inns2019 full
Vincenzo Lomonaco
 
Continual Reinforcement Learning in 3D Non-stationary Environments
Continual Reinforcement Learning in 3D Non-stationary EnvironmentsContinual Reinforcement Learning in 3D Non-stationary Environments
Continual Reinforcement Learning in 3D Non-stationary Environments
Vincenzo Lomonaco
 
Continual/Lifelong Learning with Deep Architectures
Continual/Lifelong Learning with Deep ArchitecturesContinual/Lifelong Learning with Deep Architectures
Continual/Lifelong Learning with Deep Architectures
Vincenzo Lomonaco
 
Continual Learning for Robotics
Continual Learning for RoboticsContinual Learning for Robotics
Continual Learning for Robotics
Vincenzo Lomonaco
 
Open-Source Frameworks for Deep Learning: an Overview
Open-Source Frameworks for Deep Learning: an OverviewOpen-Source Frameworks for Deep Learning: an Overview
Open-Source Frameworks for Deep Learning: an Overview
Vincenzo Lomonaco
 
Continual Learning with Deep Architectures Workshop @ Computer VISIONers Conf...
Continual Learning with Deep Architectures Workshop @ Computer VISIONers Conf...Continual Learning with Deep Architectures Workshop @ Computer VISIONers Conf...
Continual Learning with Deep Architectures Workshop @ Computer VISIONers Conf...
Vincenzo Lomonaco
 
CORe50: a New Dataset and Benchmark for Continual Learning and Object Recogni...
CORe50: a New Dataset and Benchmark for Continual Learning and Object Recogni...CORe50: a New Dataset and Benchmark for Continual Learning and Object Recogni...
CORe50: a New Dataset and Benchmark for Continual Learning and Object Recogni...
Vincenzo Lomonaco
 
Continuous Learning with Deep Architectures
Continuous Learning with Deep ArchitecturesContinuous Learning with Deep Architectures
Continuous Learning with Deep Architectures
Vincenzo Lomonaco
 
CORe50: a New Dataset and Benchmark for Continuous Object Recognition Poster
CORe50: a New Dataset and Benchmark for Continuous Object Recognition PosterCORe50: a New Dataset and Benchmark for Continuous Object Recognition Poster
CORe50: a New Dataset and Benchmark for Continuous Object Recognition Poster
Vincenzo Lomonaco
 
Continuous Unsupervised Training of Deep Architectures
Continuous Unsupervised Training of Deep ArchitecturesContinuous Unsupervised Training of Deep Architectures
Continuous Unsupervised Training of Deep Architectures
Vincenzo Lomonaco
 
Comparing Incremental Learning Strategies for Convolutional Neural Networks
Comparing Incremental Learning Strategies for Convolutional Neural NetworksComparing Incremental Learning Strategies for Convolutional Neural Networks
Comparing Incremental Learning Strategies for Convolutional Neural Networks
Vincenzo Lomonaco
 
Deep Learning for Computer Vision: A comparision between Convolutional Neural...
Deep Learning for Computer Vision: A comparision between Convolutional Neural...Deep Learning for Computer Vision: A comparision between Convolutional Neural...
Deep Learning for Computer Vision: A comparision between Convolutional Neural...
Vincenzo Lomonaco
 
Deep Learning for Computer Vision: A comparision between Convolutional Neural...
Deep Learning for Computer Vision: A comparision between Convolutional Neural...Deep Learning for Computer Vision: A comparision between Convolutional Neural...
Deep Learning for Computer Vision: A comparision between Convolutional Neural...
Vincenzo Lomonaco
 
A Framework for Deadlock Detection in Java
A Framework for Deadlock Detection in JavaA Framework for Deadlock Detection in Java
A Framework for Deadlock Detection in Java
Vincenzo Lomonaco
 
Deep Learning libraries and first experiments with Theano
Deep Learning libraries and first experiments with TheanoDeep Learning libraries and first experiments with Theano
Deep Learning libraries and first experiments with Theano
Vincenzo Lomonaco
 
Word2vec on the italian language: first experiments
Word2vec on the italian language: first experimentsWord2vec on the italian language: first experiments
Word2vec on the italian language: first experiments
Vincenzo Lomonaco
 

More from Vincenzo Lomonaco (20)

2023-08-22 CoLLAs Tutorial - Beyond CIL.pdf
2023-08-22 CoLLAs Tutorial - Beyond CIL.pdf2023-08-22 CoLLAs Tutorial - Beyond CIL.pdf
2023-08-22 CoLLAs Tutorial - Beyond CIL.pdf
 
Continual Learning with Deep Architectures - Tutorial ICML 2021
Continual Learning with Deep Architectures - Tutorial ICML 2021Continual Learning with Deep Architectures - Tutorial ICML 2021
Continual Learning with Deep Architectures - Tutorial ICML 2021
 
Toward Continual Learning on the Edge
Toward Continual Learning on the EdgeToward Continual Learning on the Edge
Toward Continual Learning on the Edge
 
Continual Learning: Another Step Towards Truly Intelligent Machines
Continual Learning: Another Step Towards Truly Intelligent MachinesContinual Learning: Another Step Towards Truly Intelligent Machines
Continual Learning: Another Step Towards Truly Intelligent Machines
 
Tutorial inns2019 full
Tutorial inns2019 fullTutorial inns2019 full
Tutorial inns2019 full
 
Continual Reinforcement Learning in 3D Non-stationary Environments
Continual Reinforcement Learning in 3D Non-stationary EnvironmentsContinual Reinforcement Learning in 3D Non-stationary Environments
Continual Reinforcement Learning in 3D Non-stationary Environments
 
Continual/Lifelong Learning with Deep Architectures
Continual/Lifelong Learning with Deep ArchitecturesContinual/Lifelong Learning with Deep Architectures
Continual/Lifelong Learning with Deep Architectures
 
Continual Learning for Robotics
Continual Learning for RoboticsContinual Learning for Robotics
Continual Learning for Robotics
 
Open-Source Frameworks for Deep Learning: an Overview
Open-Source Frameworks for Deep Learning: an OverviewOpen-Source Frameworks for Deep Learning: an Overview
Open-Source Frameworks for Deep Learning: an Overview
 
Continual Learning with Deep Architectures Workshop @ Computer VISIONers Conf...
Continual Learning with Deep Architectures Workshop @ Computer VISIONers Conf...Continual Learning with Deep Architectures Workshop @ Computer VISIONers Conf...
Continual Learning with Deep Architectures Workshop @ Computer VISIONers Conf...
 
CORe50: a New Dataset and Benchmark for Continual Learning and Object Recogni...
CORe50: a New Dataset and Benchmark for Continual Learning and Object Recogni...CORe50: a New Dataset and Benchmark for Continual Learning and Object Recogni...
CORe50: a New Dataset and Benchmark for Continual Learning and Object Recogni...
 
Continuous Learning with Deep Architectures
Continuous Learning with Deep ArchitecturesContinuous Learning with Deep Architectures
Continuous Learning with Deep Architectures
 
CORe50: a New Dataset and Benchmark for Continuous Object Recognition Poster
CORe50: a New Dataset and Benchmark for Continuous Object Recognition PosterCORe50: a New Dataset and Benchmark for Continuous Object Recognition Poster
CORe50: a New Dataset and Benchmark for Continuous Object Recognition Poster
 
Continuous Unsupervised Training of Deep Architectures
Continuous Unsupervised Training of Deep ArchitecturesContinuous Unsupervised Training of Deep Architectures
Continuous Unsupervised Training of Deep Architectures
 
Comparing Incremental Learning Strategies for Convolutional Neural Networks
Comparing Incremental Learning Strategies for Convolutional Neural NetworksComparing Incremental Learning Strategies for Convolutional Neural Networks
Comparing Incremental Learning Strategies for Convolutional Neural Networks
 
Deep Learning for Computer Vision: A comparision between Convolutional Neural...
Deep Learning for Computer Vision: A comparision between Convolutional Neural...Deep Learning for Computer Vision: A comparision between Convolutional Neural...
Deep Learning for Computer Vision: A comparision between Convolutional Neural...
 
Deep Learning for Computer Vision: A comparision between Convolutional Neural...
Deep Learning for Computer Vision: A comparision between Convolutional Neural...Deep Learning for Computer Vision: A comparision between Convolutional Neural...
Deep Learning for Computer Vision: A comparision between Convolutional Neural...
 
A Framework for Deadlock Detection in Java
A Framework for Deadlock Detection in JavaA Framework for Deadlock Detection in Java
A Framework for Deadlock Detection in Java
 
Deep Learning libraries and first experiments with Theano
Deep Learning libraries and first experiments with TheanoDeep Learning libraries and first experiments with Theano
Deep Learning libraries and first experiments with Theano
 
Word2vec on the italian language: first experiments
Word2vec on the italian language: first experimentsWord2vec on the italian language: first experiments
Word2vec on the italian language: first experiments
 

Recently uploaded

一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单
一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单
一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单
ewymefz
 
一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单
一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单
一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单
vcaxypu
 
Investigate & Recover / StarCompliance.io / Crypto_Crimes
Investigate & Recover / StarCompliance.io / Crypto_CrimesInvestigate & Recover / StarCompliance.io / Crypto_Crimes
Investigate & Recover / StarCompliance.io / Crypto_Crimes
StarCompliance.io
 
一比一原版(UofM毕业证)明尼苏达大学毕业证成绩单
一比一原版(UofM毕业证)明尼苏达大学毕业证成绩单一比一原版(UofM毕业证)明尼苏达大学毕业证成绩单
一比一原版(UofM毕业证)明尼苏达大学毕业证成绩单
ewymefz
 
一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单
一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单
一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单
nscud
 
Opendatabay - Open Data Marketplace.pptx
Opendatabay - Open Data Marketplace.pptxOpendatabay - Open Data Marketplace.pptx
Opendatabay - Open Data Marketplace.pptx
Opendatabay
 
一比一原版(NYU毕业证)纽约大学毕业证成绩单
一比一原版(NYU毕业证)纽约大学毕业证成绩单一比一原版(NYU毕业证)纽约大学毕业证成绩单
一比一原版(NYU毕业证)纽约大学毕业证成绩单
ewymefz
 
Criminal IP - Threat Hunting Webinar.pdf
Criminal IP - Threat Hunting Webinar.pdfCriminal IP - Threat Hunting Webinar.pdf
Criminal IP - Threat Hunting Webinar.pdf
Criminal IP
 
一比一原版(ArtEZ毕业证)ArtEZ艺术学院毕业证成绩单
一比一原版(ArtEZ毕业证)ArtEZ艺术学院毕业证成绩单一比一原版(ArtEZ毕业证)ArtEZ艺术学院毕业证成绩单
一比一原版(ArtEZ毕业证)ArtEZ艺术学院毕业证成绩单
vcaxypu
 
Predicting Product Ad Campaign Performance: A Data Analysis Project Presentation
Predicting Product Ad Campaign Performance: A Data Analysis Project PresentationPredicting Product Ad Campaign Performance: A Data Analysis Project Presentation
Predicting Product Ad Campaign Performance: A Data Analysis Project Presentation
Boston Institute of Analytics
 
Criminal IP - Threat Hunting Webinar.pdf
Criminal IP - Threat Hunting Webinar.pdfCriminal IP - Threat Hunting Webinar.pdf
Criminal IP - Threat Hunting Webinar.pdf
Criminal IP
 
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Subhajit Sahu
 
Jpolillo Amazon PPC - Bid Optimization Sample
Jpolillo Amazon PPC - Bid Optimization SampleJpolillo Amazon PPC - Bid Optimization Sample
Jpolillo Amazon PPC - Bid Optimization Sample
James Polillo
 
Innovative Methods in Media and Communication Research by Sebastian Kubitschk...
Innovative Methods in Media and Communication Research by Sebastian Kubitschk...Innovative Methods in Media and Communication Research by Sebastian Kubitschk...
Innovative Methods in Media and Communication Research by Sebastian Kubitschk...
correoyaya
 
Empowering Data Analytics Ecosystem.pptx
Empowering Data Analytics Ecosystem.pptxEmpowering Data Analytics Ecosystem.pptx
Empowering Data Analytics Ecosystem.pptx
benishzehra469
 
一比一原版(CU毕业证)卡尔顿大学毕业证成绩单
一比一原版(CU毕业证)卡尔顿大学毕业证成绩单一比一原版(CU毕业证)卡尔顿大学毕业证成绩单
一比一原版(CU毕业证)卡尔顿大学毕业证成绩单
yhkoc
 
Best best suvichar in gujarati english meaning of this sentence as Silk road ...
Best best suvichar in gujarati english meaning of this sentence as Silk road ...Best best suvichar in gujarati english meaning of this sentence as Silk road ...
Best best suvichar in gujarati english meaning of this sentence as Silk road ...
AbhimanyuSinha9
 
Q1’2024 Update: MYCI’s Leap Year Rebound
Q1’2024 Update: MYCI’s Leap Year ReboundQ1’2024 Update: MYCI’s Leap Year Rebound
Q1’2024 Update: MYCI’s Leap Year Rebound
Oppotus
 
一比一原版(UVic毕业证)维多利亚大学毕业证成绩单
一比一原版(UVic毕业证)维多利亚大学毕业证成绩单一比一原版(UVic毕业证)维多利亚大学毕业证成绩单
一比一原版(UVic毕业证)维多利亚大学毕业证成绩单
ukgaet
 
一比一原版(BU毕业证)波士顿大学毕业证成绩单
一比一原版(BU毕业证)波士顿大学毕业证成绩单一比一原版(BU毕业证)波士顿大学毕业证成绩单
一比一原版(BU毕业证)波士顿大学毕业证成绩单
ewymefz
 

Recently uploaded (20)

一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单
一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单
一比一原版(IIT毕业证)伊利诺伊理工大学毕业证成绩单
 
一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单
一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单
一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单
 
Investigate & Recover / StarCompliance.io / Crypto_Crimes
Investigate & Recover / StarCompliance.io / Crypto_CrimesInvestigate & Recover / StarCompliance.io / Crypto_Crimes
Investigate & Recover / StarCompliance.io / Crypto_Crimes
 
一比一原版(UofM毕业证)明尼苏达大学毕业证成绩单
一比一原版(UofM毕业证)明尼苏达大学毕业证成绩单一比一原版(UofM毕业证)明尼苏达大学毕业证成绩单
一比一原版(UofM毕业证)明尼苏达大学毕业证成绩单
 
一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单
一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单
一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单
 
Opendatabay - Open Data Marketplace.pptx
Opendatabay - Open Data Marketplace.pptxOpendatabay - Open Data Marketplace.pptx
Opendatabay - Open Data Marketplace.pptx
 
一比一原版(NYU毕业证)纽约大学毕业证成绩单
一比一原版(NYU毕业证)纽约大学毕业证成绩单一比一原版(NYU毕业证)纽约大学毕业证成绩单
一比一原版(NYU毕业证)纽约大学毕业证成绩单
 
Criminal IP - Threat Hunting Webinar.pdf
Criminal IP - Threat Hunting Webinar.pdfCriminal IP - Threat Hunting Webinar.pdf
Criminal IP - Threat Hunting Webinar.pdf
 
一比一原版(ArtEZ毕业证)ArtEZ艺术学院毕业证成绩单
一比一原版(ArtEZ毕业证)ArtEZ艺术学院毕业证成绩单一比一原版(ArtEZ毕业证)ArtEZ艺术学院毕业证成绩单
一比一原版(ArtEZ毕业证)ArtEZ艺术学院毕业证成绩单
 
Predicting Product Ad Campaign Performance: A Data Analysis Project Presentation
Predicting Product Ad Campaign Performance: A Data Analysis Project PresentationPredicting Product Ad Campaign Performance: A Data Analysis Project Presentation
Predicting Product Ad Campaign Performance: A Data Analysis Project Presentation
 
Criminal IP - Threat Hunting Webinar.pdf
Criminal IP - Threat Hunting Webinar.pdfCriminal IP - Threat Hunting Webinar.pdf
Criminal IP - Threat Hunting Webinar.pdf
 
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
 
Jpolillo Amazon PPC - Bid Optimization Sample
Jpolillo Amazon PPC - Bid Optimization SampleJpolillo Amazon PPC - Bid Optimization Sample
Jpolillo Amazon PPC - Bid Optimization Sample
 
Innovative Methods in Media and Communication Research by Sebastian Kubitschk...
Innovative Methods in Media and Communication Research by Sebastian Kubitschk...Innovative Methods in Media and Communication Research by Sebastian Kubitschk...
Innovative Methods in Media and Communication Research by Sebastian Kubitschk...
 
Empowering Data Analytics Ecosystem.pptx
Empowering Data Analytics Ecosystem.pptxEmpowering Data Analytics Ecosystem.pptx
Empowering Data Analytics Ecosystem.pptx
 
一比一原版(CU毕业证)卡尔顿大学毕业证成绩单
一比一原版(CU毕业证)卡尔顿大学毕业证成绩单一比一原版(CU毕业证)卡尔顿大学毕业证成绩单
一比一原版(CU毕业证)卡尔顿大学毕业证成绩单
 
Best best suvichar in gujarati english meaning of this sentence as Silk road ...
Best best suvichar in gujarati english meaning of this sentence as Silk road ...Best best suvichar in gujarati english meaning of this sentence as Silk road ...
Best best suvichar in gujarati english meaning of this sentence as Silk road ...
 
Q1’2024 Update: MYCI’s Leap Year Rebound
Q1’2024 Update: MYCI’s Leap Year ReboundQ1’2024 Update: MYCI’s Leap Year Rebound
Q1’2024 Update: MYCI’s Leap Year Rebound
 
一比一原版(UVic毕业证)维多利亚大学毕业证成绩单
一比一原版(UVic毕业证)维多利亚大学毕业证成绩单一比一原版(UVic毕业证)维多利亚大学毕业证成绩单
一比一原版(UVic毕业证)维多利亚大学毕业证成绩单
 
一比一原版(BU毕业证)波士顿大学毕业证成绩单
一比一原版(BU毕业证)波士顿大学毕业证成绩单一比一原版(BU毕业证)波士顿大学毕业证成绩单
一比一原版(BU毕业证)波士顿大学毕业证成绩单
 

Don't forget, there is more than forgetting: new metrics for Continual Learning - Poster

  • 1. , Don’t forget, there is more than forgetting: new metrics for Continual Learning Natalia D´ıaz-Rodr´ıguez1∗ , Vincenzo Lomonaco2∗ , Davide Maltoni2 and David Filliat1 , (1) U2IS, ENSTA ParisTech, Inria FLOWERS team, Palaiseau, France. https://flowers.inria.fr/ {natalia.diaz, david.filliat}@ensta-paristech.fr. (2) University of Bologna, Italy. {vincenzo.lomonaco, davide.maltoni}@unibo.it http://continualai.org/ (*) Both authors contributed equally to this work. , Motivation The lack of consensus in evaluating continual learning (CL) algorithms and the almost exclusive focus on catastrophic forgetting motivate us to propose a more comprehensive set of implementation independent metrics accounting for factors we believe have practical implications worth considering with respect to “static” machine learning settings. New metrics for Continual Learning (CL) A CL algorithm ACL is an algorithm with the following signature: ∀Di ∈ D, ACL i : < hi−1, Tri, Mi−1, t >→< hi, Mi > (1) Where hi is the model, Tri is the training set of examples drawn from the respective Di distribution, Mi is an external memory where we can store previous training examples and t is a task label. For simplicity, we can assume N as the number of tasks, one for each Tri. Accuracy (A): Given matrix R ∈ RN×N , which contains in each entry Ri,j the test classification accuracy of the model on task tj after observing the last sample from task ti [3], Accuracy metric is: A = N i≥j Ri,j N(N+1) 2 (2) Table 1: Accuracy matrix R: elements accounted to compute A (white & cyan), BWT (cyan), and FWT (gray) Tri = training, Tei= test tasks. R Te1 Te2 Te3 Tr1 R∗ Rij Rij Tr2 Rij R∗ Rij Tr3 Rij Rij R∗ Backward Transfer (BWT) measures the influence that learning a task has on the performance on previous tasks [3]: BWT = N i=2 i−1 j=1(Ri,j − Rj,j) N(N−1) 2 (3) Remembering (REM): REM = 1 − |min(BWT, 0)| (4) is the originally negative BWT while the originally positive BWT, i.e., improvement over time, is: Positive Backward Transfer (BWT+ ): BWT+ = max(BWT, 0) (5) Forward Transfer (FWT): measures the influence that learning a task has on the performance of future tasks [3]. FWT = N i<j Ri,j N(N−1) 2 (6) Model size efficiency (MS): The memory size of model hi quantified in terms of parameters θ at each task i, Mem(θi), should not grow too rapidly with respect to the size of the model that learned the first task, Mem(θ1). MS = min(1, N i=1 Mem(θ1) Mem(θi) N ) (7) Samples storage size efficiency (SSS): The memory occupation in bits by the samples storage memory M, Mem(M), should be bounded by the memory occupation of the total nr of examples encountered at the end of the last task (D is the lifetime dataset associated to all distributions D): SSS = 1 − min(1, N i=1 Mem(Mi) Mem(D) N ) (8) Computational efficiency (CE): it is bounded by the nr of operations for training set Tri. Ops(Tri) is the number of (mul-adds) operations needed to learn Tri and Ops ↑↓(Tri) is the nr of operations required to do one forward and one backward (backprop) pass on Tri). CE = min(1, N i=1 Ops↑↓(Tri)·ε 1+Ops(Tri) N ) (9) CLscore and CLstability aggregating metrics CLscore: if ci ∈ [0, 1]) is the avg. of r runs assigned a weight wi ∈ [0, 1] s.t. C i wi = 1: CLscore = #C i=1 wici (10) CLstability: the average of the std. devs. from all previous criteria ci: CLstability = 1 − #C i=1 wistddev(ci) (11) Experiments and Conclusion Figure 1: a) Spider chart: CL metrics per strategy (larger area is better). b) Accuracy per CL strategy computed over the fixed test set as proposed in [4]. We evaluate the CL metrics on cumulative and na¨ıve baseline strategies as in [4], Elastic Weight Consolidation (EWC) [1], Synaptic Intelligence (SI) [5] and Learning without Forgetting (LwF) [2] on iCIFAR-100 dataset. Table 2: CL metrics and CLscore for each CL strategy evaluated (higher is better). Str. A REM BWT+ FWT MS SSS CE CLsco CLstab Na¨ı 0.3825 0.6664 0.00 0.10 1.00 1.00 0.4492 0.5140 0.9986 Cum 0.7225 1.00 0.0673 0.10 1.00 0.55 0.1496 0.5128 0.9979 EWC 0.5940 0.9821 0.00 0.10 0.40 1.00 0.3495 0.4894 0.9972 LWF 0.5278 0.9667 0.00 0.10 1.00 1.00 0.4429 0.5768 0.9986 SI 0.5795 0.9620 0.00 0.10 0.40 1.00 0.3613 0.4861 0.9970 • Experiments: 3 weight configs W =[wA, wMS, wSSS, wCE, wREM, wBWT+, wFWT ] (here wi = 1 #C, CNN in [5, 4]). • Future work: provide insights that assess the importance of different metric schemes and their entanglement, and how to use these metrics wisely to assist choosing among algorithms. References [1] J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, D. Hassabis, C. Clopath, D. Kumaran, and R. Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521–3526, 2017. [2] Z. Li and D. Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935–2947, Dec 2018. [3] D. Lopez-Paz et al. Gradient episodic memory for continual learning. In Advances in Neural Information Processing Systems, pages 6467–6476, 2017. [4] D. Maltoni and V. Lomonaco. Continuous learning in single-incremental-task scenarios. arXiv preprint arXiv:1806.08568, 2018. [5] F. Zenke, B. Poole, and S. Ganguli. Continual learning through synaptic intelligence. In Proceedings of the 34th International Conference on Machine Learning, Proceedings of Machine Learning Research, pages 3987–3995. PMLR, 2017.