SlideShare uma empresa Scribd logo
1 de 47
Continuous Unsupervised Training
of Deep Architectures
Biometric System Laboratory
DISI - University of Bologna
http://biolab.csr.unibo.it
6th July 2017 – IIT Genova
Davide Maltoni
davide.maltoni@unibo.it
Vincenzo Lomonaco
vincenzo.lomonaco@unibo.it
1Biometric System Laboratory
Outline
Deep architectures
Supervised, Semi supervised and Unsupervised learning
Continuous Incremental learning & Forgetting
Temporal Coherence for Semi Supervised Learning
Core50: Dataset & Benchmark
2Biometric System Laboratory
Deep architectures  Hierarchical architectures with many layers
Neural Networks (with more than 3 levels)
Convolutional Neural Networks (CNN)
Hierarchical Temporal Memory (HTM)
Multi-stage Hubel-Wiesel Architectures (Ranzato 2007)
Alternating layers of feature detectors (Hubel and Wiesel’s simple cells)
and local pooling/subsampling of features (Hubel and Wiesel’s complex
cells).
We focused on CNN and HTM
Increasing the number of alternating feature extraction (discrimination)
and feature pooling (generalization) layers seems to be the key to tackle
invariance
Deep architectures (1)
3Biometric System Laboratory
State-of-the-art achievements in many fields:
vision (object classification, detection, segmentation)
speech recognition
natural language processing
Mainly supervised training (backpropagation like) with huge and
fixed datasets:
ImageNet - ILSVRC (1.2M labeled images - 1,000 classes)
YouTube-8M (7M videos – 4,716 classes)
Microsoft Coco (2M instances, 300,000 images – 80 classes)
GPU parallelism to speed-up training (months  days)
Deep architectures (2)
4Biometric System Laboratory
Learning features
Learning Features by Watching Objects Move (Pathak et al, 2017)
Unsupervised motion-based segmentation as pseudo ground truth
Sparse Autoencoder objective function designed to optimize the input
reconstruction (promote sparsity and feature grouping)
Learning low level feature detectors (V1 like) is quite simple (Olshausen & Field,
1996), but learning high level feature detectors requires huge amount of data (Le
2012, Google)
Unsupervised learning
reconstruction pooling
5Biometric System Laboratory
Semi-supervised
Typically a small set of labeled data + large set of unlabeled data. Knowledge
of pattern absolute density can help to optimize the decision boundary
Self-training
a system is first trained with a small amount of labeled data and then used to
classify the unlabeled data. The most confident unlabeled points are
(iteratively) added to the training set.
but unfortunately:
patterns whose label can be correctly guessed do not bring much value to
improve the current representation
really useful patterns (in term of diversity) are not added because of the low self-
confidence
Semi-supervised training
(Xiaojin Zhu, 2007)
6Biometric System Laboratory
Adapt an already trained architecture to solve a new problem (in
the same domain)
Very popular today (avoid time consuming retraining from scratch)
Replace (at least) the last layer (top classifier)
Reuse of features
– Keep them fixed and train only top layers
– Tune them together with top layers
Is not incremental
– Relevant accuracy degradation on the original task
Transfer learning
7Biometric System Laboratory
Why it is important:
Many real-world scenarios are not fixed a priori
Training data often available as subsequent batches or from streaming sources
Main Strategies:
store the past data and retrain the system from scratch (often unrealistic)
data comes in sequential batches, used once and then lost (more feasible)
Problems:
– Catastrophic forgetting (McCloskey & Cohen 1989)
– Stability-plasticity dilemma (Mermillod et al. 2013)
Continuous - Incremental learning
8Biometric System Laboratory
Solutions:
– Incremental Classifiers – Camoriano et. al. 2016, Lomonaco & Maltoni 2016*
Keep feature extraction fixed and train an incremental classifier (e.g. RLS, SVM)
– Early stopping
each batch produces only slight changes in the parameters
– Self-refreshing
generative models to create pseudo-patterns
– Elastic Weight Consolidation (EWC) – Kirkpatrick et al. 2016
move only non-critical weights
– Learning without Forgetting (LwF) – Li & Hoiem 2016
promote output stability (of the old network on new patterns)
Interesting approaches, but tested only on simple (two or few task) problems.
Continuous - Incremental learning (2)
* Lomonaco and Maltoni, Comparing Incremental Learning Strategies for Convolutional Neural Networks, Workshop on
Artificial Neural Networks in Pattern Recognition, 2016.
9Biometric System Laboratory
Natural learning
Continuous / Lifelong (and possibly online)
Partially supervised (or with reinforcement), but mostly unsupervised
Multimodal / Multitask
human-like learning involves an initial small amount of direct instruction (e.g.
parental labeling of objects during childhood) combined with large amounts of
subsequence unsupervised experience (e.g. self-interaction with objects)
Ultimate interest
10Biometric System Laboratory
HTM
Hierarchical Temporal Memory (HTM) is a biologically-inspired
computational framework proposed by Hawkins and George
Dileep George, Jeff Hawkins, “A Hierarchical Bayesian Model of Invariant Pattern
Recognition in the Visual Cortex”, IJCNN 2005.
Dileep George, Jeff Hawkins, “Towards a Mathematical Theory of Cortical Micro-
circuits”, PLoS Comput. Biol. 5(10), 2009.
A silicon-valley company (Numenta) created to develop/market HTM technology.
HTM review and implementation details in:
Davide Maltoni, “Pattern recognition by Hierarchical Temporal Memory”, Tech.
Report, DEIS – University of Bologna, April 2011:
http://cogprints.org/9187/1/HTM_TR_v1.0.pdf
HTM Supervised Refinement (HSR) is a backpropagation-like approach
to tune HTM after initial Pre-training:
Rehn, E. M. and Maltoni, D. “Incremental learning by message passing in hierarchical
temporal memory”. Neural Computation, 26(8):1763–1809, 2014.
11Biometric System Laboratory
HTM generations
We work on the “first generation” of HTM.
In 2011 Numenta focused on a different model (called CLA: Cortical
Learning Algorithms) :
better suited for dynamic patterns and their temporal relations (compete with
Recurrent NN, LSTM).
Practical applications of CLA are tools for anomaly detection on data streams (e.g.
stock market data, network intrusion detection, etc.).
CLA is not efficient enough to work with high-dimensional patterns such as images.
12Biometric System Laboratory
HTM Architecture
Tree-like network with
nlevs (≥ 2) levels:
Level 3
(output)
1 node
Level 2
(intermediate)
2×2 nodes
Level 1
(intermediate)
4×4 nodes
Level 0
(input)
16×16 nodes
Image
16×16 pixels
Example: 4 levels,
16x16 input
13Biometric System Laboratory
Node structure
Input nodes (level 0) are simple interfaces between pixels and level 1 nodes.
Intermediate nodes are the real computation engine of HTM: they store
coincidences and temporal groups.
A single output node works as a pattern classifiers (e.g., a NN classifiers).
𝛌2
−
…𝛌1
− 𝛌 𝑚
−
Coincidences C
Temporal groups GMatrix
𝐜2 …𝐜1 𝐜 𝑛 𝑐
𝐠2 …𝐠1 𝐠 𝑛 𝑔𝐏𝐂𝐆
𝛌−
= 𝛌1
−
, 𝛌2
−
… 𝛌 𝑚
−
𝐲
𝛌+
𝛌2
−
…𝛌1
− 𝛌 𝑚
−
Coincidences C
Prior class prob.Matrix
𝐜2 …𝐜1 𝐜 𝑛 𝑐
[𝑃(𝑤1), 𝑃(𝑤2) … 𝑃(𝑤𝑛 𝑤
)]𝐏𝐂𝐖
𝛌−
= 𝛌1
−
, 𝛌2
−
… 𝛌 𝑚
−
𝐲
[𝑃(𝑤1|𝑒), 𝑃(𝑤2|𝑒) … 𝑃(𝑤𝑛 𝑤
|𝑒)]
Intermediate
node
Output
node
14Biometric System Laboratory
Coincidences
Each coincidence is a sort of feature extractor that spans a portion of the image
corresponding to the node receptive field (i.e., small at low levels and large at high
levels). Coincidences are used to perform a spatial analysis of input patterns and to
find out spatial similarities.
Example of level 1 coincidences
15Biometric System Laboratory
Temporal groups
A temporal group is a subset of coincidences, that could be spatially quite different
each from the other, but that are likely to be originated from simple variations of the
same pattern.
HTM exploits temporal smoothness to create temporal groups: patterns presented
to the network very close in time, are likely to be variants of the same pattern that is
smoothly moving throughout the network field of view.
Example of level 1 groups
𝐠1
𝐠2
𝐠3
𝐠4
𝐠5
𝐠6
𝐠7
𝐠8
𝐠9
𝐠10
𝐠11
𝐠12
𝐠13
𝐠14
16Biometric System Laboratory
HTM vs CNN
Key points:
Bayesian probabilistic formulation: Bayesian Network + Bayesian Belief propagation
equations. Parameters values constrained by probability laws.
Top down and bottom-up information flow: feedback messages from higher levels
carry contextual information to bias the behavior of lower levels. By fusing bottom up
and top down messages each HTM node reaches an internal state (called node belief
and corresponding to Bayes posterior) which is an optimal probabilistic explanation of
the external stimuli.
Pooling is a key component in HTM (some CNN work well even without pooling
layers).
Unsupervised Pre-training is very effective in HTM.
Compares favorably on small-scale problems, difficult to scale-up. Designing large
HTM is not easy (architectural choices, pre-training, code optimization, numerical
problems).
16x16  32x32  64x64  128x128  …
17Biometric System Laboratory
From supervised to Semi-Supervised Tuning
Exploiting time (again) as supervisor.
Biological plausibility: … hypothesis under which invariance is learned
from temporal continuity of object features during natural visual
experience without external supervision (Li & DiCarlo 2008)
SST idea: pass back the current output vector as desired vector (i.e.,
label) for the next pattern
D. Maltoni and V. Lomonaco, Semi-supervised Tuning from Temporal Coherence,
International Conference on Pattern Recognition, 2016.
ArXiv 1511.03163
18Biometric System Laboratory
SST (1)
Temporal coherent patterns: 𝒗 𝑡 , 𝑡 = 1 … 𝑚
A classifier 𝑵 maps an input pattern to an output vector: 𝒗 𝑡 → 𝑵 𝒗 𝑡
Loss function to be minimized (e.g. backprop):
1
2
𝑵 𝒗 𝑡 − 𝒅(𝒗 𝑡 )
2
1. Supervised Tuning (SupT):
𝒅 𝒗 𝑡
= 𝚫 𝑤 = 0, … , 1, … 0 w is the pattern class
2. Supervised Tuning with Regularization (SupTR):
𝒅 𝒗 𝑡
= 𝜆 ∙ 𝚫 𝑤 + 1 − 𝜆 ∙ 𝑵 𝒗 𝑡−1
position w
19Biometric System Laboratory
SST (2)
3. Semi Supervised Tuning – Basic (SST-B):
𝒅 𝒗 𝑡 = 𝑵 𝒗 𝑡−1
4. Semi Supervised Tuning – Advanced (SST-A):
𝒇 𝒗 𝑡
=
𝑵 𝒗 𝑡−1 𝑡 = 2
𝒇 𝒗 𝑡−1 + 𝑵 𝒗 𝑡−1
2
𝑡 > 2
𝒅 𝒗 𝑡 =
𝒇 𝒗 𝑡 𝑖𝑓 max
𝑖
𝒇𝑖 𝒗 𝑡 > 𝑠𝑐
𝑵 𝒗 𝑡 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
20Biometric System Laboratory
SST Experiments
We generated temporal coherent sequences from NORB and COIL100
datasets by randomly walking the variation space (pose, lighting).
1 sequence for initial (supervised training) + 9 sequences for
incremental tuning
Test set contains frames with a given mindist from all training patterns
21Biometric System Laboratory
SST Experiments (2)
HTM on NORB CNN on NORB
22Biometric System Laboratory
SST Experiments (3)
making the problem harder by
reducing initial accuracy …
Comparison with
self-training
Self-training
23Biometric System Laboratory
Ongoing work on SST
SST approach very effective with HTM
SST-A accuracy is sometime very close to supervised scenario.
Why does HTM work better than CNN with SST?
HTM seems to be more robust w.r.t. forgetting, Why?
In HTM we keeps coincidence fixed. Try with CNN - NiN (Network in Network) models
where we move only feature pooling weights.
Investigate other architectural differences.
New experiments on natural datasets and continuous learning
scenarios.
Core50
iCubWorld-Transf
24Biometric System Laboratory
CORe50
Dataset, Benchmark, code and additional information freely
available at: https://vlomonaco.github.io/core50
25Biometric System Laboratory
CORe50 (2)
Motivation
Temporal Coherence has already shown to be a good surrogate supervised signal
(also in complex sequence learning problems)
Still, it’s very difficult to find simple videos of objects smoothly moving in front of the
camera.
Moreover, in order to assess Continuous Learning scenarios we need the presence of
multiple (temporal coherent and unconstrained) views of the same objects taken in
different sessions (varying background, lighting, pose, occlusions, etc.)
Sometimes it is possible to generate a number of exploration sequences that turn a
native static benchmarks into continuous learning tasks, but the result is often
unnatural (Maltoni & Lomonaco 2016)
26Biometric System Laboratory
Comparison with other Datasets
27Biometric System Laboratory
50 Objects of 10 Classes
28Biometric System Laboratory
11 Sessions
One frame of the same object (#41) throughout the 11 acquisition sessions.
Three of the eleven sessions (#3, #7 and #10) have been selected for test and
the remaining 8 sessions are used for training.
29Biometric System Laboratory
Acquisition Setting
For each session and for each object, a 15 seconds video (at 20 fps) has been
recorded with a Kinect 2.0 sensor delivering 300 RGB-D frames.
The acquisition interface identifies a central region where the object should be
this allows to crop the frame size from 1024×575 to 350×350.
30Biometric System Laboratory
Acquisition setting (2)
The 350×350 frames are then automatically cropped to 128×128 based on a
fully automated tracker.
Example of 1 second recording (at 20 fps) of object #26 in session #4 (outdoor).
Note the smooth movement, pose change and partial occlusion.
31Biometric System Laboratory
Continuous Learning Scenarios
New Instances (NI)
New training patterns of the same classes become available in subsequent
batches with new poses and conditions (illumination, background, occlusion,
etc.).
New Classes (NC)
New training patterns belonging to different classes become available in
subsequent batches.
New Instances and Classes (NIC)
New training patterns belonging both to known and new classes become
available in subsequent training batches.
32Biometric System Laboratory
Continuous Learning Strategies
Cumulative (non-continuous)
We re-train the entire model from scratch as soon as a new batch of data is
available.
Naïve
We simply continue back-propagation with early-stopping and low learning
rate.
Copy Weights with Re-init (CWR)
Simple baseline to add new classes easily disentangling the weights
effecting each class.
33Biometric System Laboratory
Copy Weights with Re-init (CWR)
In CWR we skip layers fc6 and fc7 and directly connect pool5 to a final layer fc8
(followed by softmax) while maintaining the weights up to Pool5 fixed.
This allows isolating the subsets of weights that each class uses.
During the training two sets of weights are maintained by the model for the pool5→fc8
connections: cw are the consolidated weights used for inference and tw the temporary
weights used for training.
After the training of each batch temporary weights tw are copied in cw.
twcwfixed
34Biometric System Laboratory
Copy Weights with Re-init (CWR)
In CWR we skip layers fc6 and fc7 and directly connect pool5 to a final layer fc8
(followed by softmax) while maintaining the weights up to Pool5 fixed.
This allows isolating the subsets of weights that each class uses.
During the training two sets of weights are maintained by the model for the pool5→fc8
connections: cw are the consolidated weights used for inference and tw the temporary
weights used for training.
After the training of each batch temporary weights tw are copied in cw.
twcwfixed
35Biometric System Laboratory
Copy Weights with Re-init (CWR)
In CWR we skip layers fc6 and fc7 and directly connect pool5 to a final layer fc8
(followed by softmax) while maintaining the weights up to Pool5 fixed.
This allows isolating the subsets of weights that each class uses.
During the training two sets of weights are maintained by the model for the pool5→fc8
connections: cw are the consolidated weights used for inference and tw the temporary
weights used for training.
After the training of each batch temporary weights tw are copied in cw.
twcwfixed
36Biometric System Laboratory
Copy Weights with Re-init (CWR)
In CWR we skip layers fc6 and fc7 and directly connect pool5 to a final layer fc8
(followed by softmax) while maintaining the weights up to Pool5 fixed.
This allows isolating the subsets of weights that each class uses.
During the training two sets of weights are maintained by the model for the pool5→fc8
connections: cw are the consolidated weights used for inference and tw the temporary
weights used for training.
After the training of each batch temporary weights tw are copied in cw.
twcwfixed
37Biometric System Laboratory
Copy Weights with Re-init (CWR)
In CWR we skip layers fc6 and fc7 and directly connect pool5 to a final layer fc8
(followed by softmax) while maintaining the weights up to Pool5 fixed.
This allows isolating the subsets of weights that each class uses.
During the training two sets of weights are maintained by the model for the pool5→fc8
connections: cw are the consolidated weights used for inference and tw the temporary
weights used for training.
After the training of each batch temporary weights tw are copied in cw.
twcwfixed
38Biometric System Laboratory
Copy Weights with Re-init (CWR)
In CWR we skip layers fc6 and fc7 and directly connect pool5 to a final layer fc8
(followed by softmax) while maintaining the weights up to Pool5 fixed.
This allows isolating the subsets of weights that each class uses.
During the training two sets of weights are maintained by the model for the pool5→fc8
connections: cw are the consolidated weights used for inference and tw the temporary
weights used for training.
After the training of each batch temporary weights tw are copied in cw.
twcwfixed
39Biometric System Laboratory
New Instances (NI)
In this scenario the training batches coincides with the 8 sessions available in
the training set.
For each of the 50 objects, training a model on the first session ad tuning it 7
times (on the remaining 7 sessions).
40Biometric System Laboratory
New Classes (NC)
Each batch contains the whole training sequences (8) of a small group of
classes, and therefore no memory refresh is possible across batches.
In the first batch we include 10 classes, while the remaining 8 batches contain 5
classes each.
41Biometric System Laboratory
New Instances and Classes (NIC)
The first batch includes 10 classes, and the subsequent batches 5 classes
each.
However, only one training sequence per class is here included in a batch, thus
resulting in a double partitioning scheme (i.e., classes and sequences). The
total number of batches is 79.
42Biometric System Laboratory
Conclusions
Continuous/Lifelong Learning enables a boarder range of real-word applications
and counts a number of different advantages
Biologically grounded and memory/computational efficient
Enables adaptation and knowledge reuse/refining
Ideal fit for unsupervised streaming perception data (Visual, Audio, etc..)
Focus on all the three main scenarios NI, NC and NIC not just NC
Recent literature on Continuous/Lifelong learning focuses on simple sequences of
different tasks (maximum 3-4).
Temporal coherence is one of the key towards unsupervised continuous
learning
Sequence learning, Semi-Supervised Tuning (SST), etc…
43Biometric System Laboratory
Future Plans
Continuing out work in SST and CORe50
HTM vs CNN comparison
Further tests on CORe50 and iCubWorld-Transf
Implementing and evaluating LwF and EWC on Core50
LwF and EWC have been only tested in a NC-like scenario.
It is not trivial to apply them in the NI and NIC scenarios.
Thank you for listening!
Biometric System Laboratory
DISI - University of Bologna
http://biolab.csr.unibo.it
6th July 2017 – IIT Genova
Davide Maltoni
davide.maltoni@unibo.it
Vincenzo Lomonaco
vincenzo.lomonaco@unibo.it
Continuous Unsupervised Training of Deep Architectures
45Biometric System Laboratory
CNN incremental (supervised) training (2)
incremental
Strategy day1 day2 day3
Lenet7 36,00% 37,55% 35,84%
AlexNet + RLS 58,20% 66,39% 74,10%
AlexNet + SVM 56,32% 66,19% 72,07%
FRCNN + Svm 51,63% 59,63% 70,00%
AlexNet + finetuning (last lvl) 54,01% 60,00% 65,38%
AlexNet + finetuning 70,70% 77,14% 76,85%
IcubWorld28
Img_size: 128x128
Inc_batches: 3
Num_classes: 5 (x4 obj)
Time
Time
46Biometric System Laboratory
CNN incremental (supervised) training (3)
Big-brother (setB)
Img_size: 70x70
Inc_batches: 56
Num_classes: 7

Mais conteúdo relacionado

Mais procurados

Unsupervised learning
Unsupervised learningUnsupervised learning
Unsupervised learningamalalhait
 
Analysis of Neocognitron of Neural Network Method in the String Recognition
Analysis of Neocognitron of Neural Network Method in the String RecognitionAnalysis of Neocognitron of Neural Network Method in the String Recognition
Analysis of Neocognitron of Neural Network Method in the String RecognitionIDES Editor
 
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 2021Vincenzo Lomonaco
 
Summary Of Thesis
Summary Of ThesisSummary Of Thesis
Summary Of Thesisguestb452d6
 
Machine Learning, LIX004M5
Machine Learning, LIX004M5Machine Learning, LIX004M5
Machine Learning, LIX004M5butest
 
Application of soft computing techniques in electrical engineering
Application of soft computing techniques in electrical engineeringApplication of soft computing techniques in electrical engineering
Application of soft computing techniques in electrical engineeringSouvik Dutta
 
Model Evaluation in the land of Deep Learning
Model Evaluation in the land of Deep LearningModel Evaluation in the land of Deep Learning
Model Evaluation in the land of Deep LearningPramit Choudhary
 
Do Computers and Humans See Similarly?
Do Computers and Humans See Similarly?Do Computers and Humans See Similarly?
Do Computers and Humans See Similarly?Zach Monge, PhD
 
How to create a mind
How to create a mindHow to create a mind
How to create a mindrupsdino1997
 
Introduction to Interpretable Machine Learning
Introduction to Interpretable Machine LearningIntroduction to Interpretable Machine Learning
Introduction to Interpretable Machine LearningNguyen Giang
 
A systematic review on sequence-to-sequence learning with neural network and ...
A systematic review on sequence-to-sequence learning with neural network and ...A systematic review on sequence-to-sequence learning with neural network and ...
A systematic review on sequence-to-sequence learning with neural network and ...IJECEIAES
 
Human in the loop: Bayesian Rules Enabling Explainable AI
Human in the loop: Bayesian Rules Enabling Explainable AIHuman in the loop: Bayesian Rules Enabling Explainable AI
Human in the loop: Bayesian Rules Enabling Explainable AIPramit Choudhary
 
Kernel analysis of deep networks
Kernel analysis of deep networksKernel analysis of deep networks
Kernel analysis of deep networksBehrang Mehrparvar
 
Soft Computing
Soft ComputingSoft Computing
Soft ComputingMANISH T I
 
A Time Series ANN Approach for Weather Forecasting
A Time Series ANN Approach for Weather ForecastingA Time Series ANN Approach for Weather Forecasting
A Time Series ANN Approach for Weather Forecastingijctcm
 
Student intervention detection using deep learning technique
Student intervention detection using deep learning techniqueStudent intervention detection using deep learning technique
Student intervention detection using deep learning techniqueVenkat Projects
 
A Beginner’S Guide To Simulation In Immunology
A Beginner’S Guide To Simulation In ImmunologyA Beginner’S Guide To Simulation In Immunology
A Beginner’S Guide To Simulation In Immunologygpfigueredo
 
A survey on methods and applications of meta-learning with GNNs
A survey on methods and applications of meta-learning with GNNsA survey on methods and applications of meta-learning with GNNs
A survey on methods and applications of meta-learning with GNNsShreya Goyal
 

Mais procurados (20)

Unsupervised learning
Unsupervised learningUnsupervised learning
Unsupervised learning
 
Analysis of Neocognitron of Neural Network Method in the String Recognition
Analysis of Neocognitron of Neural Network Method in the String RecognitionAnalysis of Neocognitron of Neural Network Method in the String Recognition
Analysis of Neocognitron of Neural Network Method in the String Recognition
 
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
 
Summary Of Thesis
Summary Of ThesisSummary Of Thesis
Summary Of Thesis
 
Machine Learning, LIX004M5
Machine Learning, LIX004M5Machine Learning, LIX004M5
Machine Learning, LIX004M5
 
Application of soft computing techniques in electrical engineering
Application of soft computing techniques in electrical engineeringApplication of soft computing techniques in electrical engineering
Application of soft computing techniques in electrical engineering
 
Model Evaluation in the land of Deep Learning
Model Evaluation in the land of Deep LearningModel Evaluation in the land of Deep Learning
Model Evaluation in the land of Deep Learning
 
Do Computers and Humans See Similarly?
Do Computers and Humans See Similarly?Do Computers and Humans See Similarly?
Do Computers and Humans See Similarly?
 
How to create a mind
How to create a mindHow to create a mind
How to create a mind
 
MaLAI_Hyderabad presentation
MaLAI_Hyderabad presentationMaLAI_Hyderabad presentation
MaLAI_Hyderabad presentation
 
Introduction to Interpretable Machine Learning
Introduction to Interpretable Machine LearningIntroduction to Interpretable Machine Learning
Introduction to Interpretable Machine Learning
 
DEFENSE
DEFENSEDEFENSE
DEFENSE
 
A systematic review on sequence-to-sequence learning with neural network and ...
A systematic review on sequence-to-sequence learning with neural network and ...A systematic review on sequence-to-sequence learning with neural network and ...
A systematic review on sequence-to-sequence learning with neural network and ...
 
Human in the loop: Bayesian Rules Enabling Explainable AI
Human in the loop: Bayesian Rules Enabling Explainable AIHuman in the loop: Bayesian Rules Enabling Explainable AI
Human in the loop: Bayesian Rules Enabling Explainable AI
 
Kernel analysis of deep networks
Kernel analysis of deep networksKernel analysis of deep networks
Kernel analysis of deep networks
 
Soft Computing
Soft ComputingSoft Computing
Soft Computing
 
A Time Series ANN Approach for Weather Forecasting
A Time Series ANN Approach for Weather ForecastingA Time Series ANN Approach for Weather Forecasting
A Time Series ANN Approach for Weather Forecasting
 
Student intervention detection using deep learning technique
Student intervention detection using deep learning techniqueStudent intervention detection using deep learning technique
Student intervention detection using deep learning technique
 
A Beginner’S Guide To Simulation In Immunology
A Beginner’S Guide To Simulation In ImmunologyA Beginner’S Guide To Simulation In Immunology
A Beginner’S Guide To Simulation In Immunology
 
A survey on methods and applications of meta-learning with GNNs
A survey on methods and applications of meta-learning with GNNsA survey on methods and applications of meta-learning with GNNs
A survey on methods and applications of meta-learning with GNNs
 

Semelhante a Continuous Unsupervised Training of Deep Architectures

Tsinghua invited talk_zhou_xing_v2r0
Tsinghua invited talk_zhou_xing_v2r0Tsinghua invited talk_zhou_xing_v2r0
Tsinghua invited talk_zhou_xing_v2r0Joe Xing
 
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 HypeSiby Jose Plathottam
 
Summary.ppt
Summary.pptSummary.ppt
Summary.pptbutest
 
A tutorial on deep learning at icml 2013
A tutorial on deep learning at icml 2013A tutorial on deep learning at icml 2013
A tutorial on deep learning at icml 2013Philip Zheng
 
Neural Networks-introduction_with_prodecure.pptx
Neural Networks-introduction_with_prodecure.pptxNeural Networks-introduction_with_prodecure.pptx
Neural Networks-introduction_with_prodecure.pptxRatuRumana3
 
Brain Networks
Brain NetworksBrain Networks
Brain NetworksJimmy Lu
 
Brief Tour of Machine Learning
Brief Tour of Machine LearningBrief Tour of Machine Learning
Brief Tour of Machine Learningbutest
 
ppt slides
ppt slidesppt slides
ppt slidesbutest
 
NIPS2007: deep belief nets
NIPS2007: deep belief netsNIPS2007: deep belief nets
NIPS2007: deep belief netszukun
 
Machine Learning Meets Human Learning
Machine Learning Meets Human LearningMachine Learning Meets Human Learning
Machine Learning Meets Human Learningbutest
 
Deep learning for Computer Vision intro
Deep learning for Computer Vision introDeep learning for Computer Vision intro
Deep learning for Computer Vision introNadav Carmel
 
Introduction to deep learning
Introduction to deep learningIntroduction to deep learning
Introduction to deep learningAmr Rashed
 
How machines can take decisions
How machines can take decisionsHow machines can take decisions
How machines can take decisionsDeepu S Nath
 
How machines can take decisions
How machines can take decisionsHow machines can take decisions
How machines can take decisionsDeepu S Nath
 
Machine learning in science and industry — day 4
Machine learning in science and industry — day 4Machine learning in science and industry — day 4
Machine learning in science and industry — day 4arogozhnikov
 
Convolutional Neural Networks square measure terribly kind of like n.pdf
Convolutional Neural Networks square measure terribly kind of like n.pdfConvolutional Neural Networks square measure terribly kind of like n.pdf
Convolutional Neural Networks square measure terribly kind of like n.pdfpoddaranand1
 
Reservoir computing fast deep learning for sequences
Reservoir computing   fast deep learning for sequencesReservoir computing   fast deep learning for sequences
Reservoir computing fast deep learning for sequencesClaudio Gallicchio
 
Kernel methods in machine learning
Kernel methods in machine learningKernel methods in machine learning
Kernel methods in machine learningbutest
 

Semelhante a Continuous Unsupervised Training of Deep Architectures (20)

Tsinghua invited talk_zhou_xing_v2r0
Tsinghua invited talk_zhou_xing_v2r0Tsinghua invited talk_zhou_xing_v2r0
Tsinghua invited talk_zhou_xing_v2r0
 
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
 
lecun-01.ppt
lecun-01.pptlecun-01.ppt
lecun-01.ppt
 
Deep learning and computer vision
Deep learning and computer visionDeep learning and computer vision
Deep learning and computer vision
 
Summary.ppt
Summary.pptSummary.ppt
Summary.ppt
 
A tutorial on deep learning at icml 2013
A tutorial on deep learning at icml 2013A tutorial on deep learning at icml 2013
A tutorial on deep learning at icml 2013
 
Neural Networks-introduction_with_prodecure.pptx
Neural Networks-introduction_with_prodecure.pptxNeural Networks-introduction_with_prodecure.pptx
Neural Networks-introduction_with_prodecure.pptx
 
Brain Networks
Brain NetworksBrain Networks
Brain Networks
 
Brief Tour of Machine Learning
Brief Tour of Machine LearningBrief Tour of Machine Learning
Brief Tour of Machine Learning
 
ppt slides
ppt slidesppt slides
ppt slides
 
NIPS2007: deep belief nets
NIPS2007: deep belief netsNIPS2007: deep belief nets
NIPS2007: deep belief nets
 
Machine Learning Meets Human Learning
Machine Learning Meets Human LearningMachine Learning Meets Human Learning
Machine Learning Meets Human Learning
 
Deep learning for Computer Vision intro
Deep learning for Computer Vision introDeep learning for Computer Vision intro
Deep learning for Computer Vision intro
 
Introduction to deep learning
Introduction to deep learningIntroduction to deep learning
Introduction to deep learning
 
How machines can take decisions
How machines can take decisionsHow machines can take decisions
How machines can take decisions
 
How machines can take decisions
How machines can take decisionsHow machines can take decisions
How machines can take decisions
 
Machine learning in science and industry — day 4
Machine learning in science and industry — day 4Machine learning in science and industry — day 4
Machine learning in science and industry — day 4
 
Convolutional Neural Networks square measure terribly kind of like n.pdf
Convolutional Neural Networks square measure terribly kind of like n.pdfConvolutional Neural Networks square measure terribly kind of like n.pdf
Convolutional Neural Networks square measure terribly kind of like n.pdf
 
Reservoir computing fast deep learning for sequences
Reservoir computing   fast deep learning for sequencesReservoir computing   fast deep learning for sequences
Reservoir computing fast deep learning for sequences
 
Kernel methods in machine learning
Kernel methods in machine learningKernel methods in machine learning
Kernel methods in machine learning
 

Mais de 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.pdfVincenzo Lomonaco
 
Toward Continual Learning on the Edge
Toward Continual Learning on the EdgeToward Continual Learning on the Edge
Toward Continual Learning on the EdgeVincenzo 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 MachinesVincenzo 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 EnvironmentsVincenzo Lomonaco
 
Continual/Lifelong Learning with Deep Architectures
Continual/Lifelong Learning with Deep ArchitecturesContinual/Lifelong Learning with Deep Architectures
Continual/Lifelong Learning with Deep ArchitecturesVincenzo Lomonaco
 
Continual Learning for Robotics
Continual Learning for RoboticsContinual Learning for Robotics
Continual Learning for RoboticsVincenzo Lomonaco
 
Don't forget, there is more than forgetting: new metrics for Continual Learni...
Don't forget, there is more than forgetting: new metrics for Continual Learni...Don't forget, there is more than forgetting: new metrics for Continual Learni...
Don't forget, there is more than forgetting: new metrics for Continual Learni...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 OverviewVincenzo 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 ArchitecturesVincenzo 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 PosterVincenzo 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 NetworksVincenzo 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 JavaVincenzo 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 TheanoVincenzo 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 experimentsVincenzo Lomonaco
 
Machine Learning for Automated Reasoning: An Overview
Machine Learning for Automated Reasoning: An OverviewMachine Learning for Automated Reasoning: An Overview
Machine Learning for Automated Reasoning: An OverviewVincenzo Lomonaco
 

Mais de 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
 
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
 
Don't forget, there is more than forgetting: new metrics for Continual Learni...
Don't forget, there is more than forgetting: new metrics for Continual Learni...Don't forget, there is more than forgetting: new metrics for Continual Learni...
Don't forget, there is more than forgetting: new metrics for Continual Learni...
 
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
 
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
 
Machine Learning for Automated Reasoning: An Overview
Machine Learning for Automated Reasoning: An OverviewMachine Learning for Automated Reasoning: An Overview
Machine Learning for Automated Reasoning: An Overview
 

Último

GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]📊 Markus Baersch
 
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝DelhiRS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhijennyeacort
 
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...GQ Research
 
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)jennyeacort
 
Predictive Analysis for Loan Default Presentation : Data Analysis Project PPT
Predictive Analysis for Loan Default  Presentation : Data Analysis Project PPTPredictive Analysis for Loan Default  Presentation : Data Analysis Project PPT
Predictive Analysis for Loan Default Presentation : Data Analysis Project PPTBoston Institute of Analytics
 
Data Analysis Project : Targeting the Right Customers, Presentation on Bank M...
Data Analysis Project : Targeting the Right Customers, Presentation on Bank M...Data Analysis Project : Targeting the Right Customers, Presentation on Bank M...
Data Analysis Project : Targeting the Right Customers, Presentation on Bank M...Boston Institute of Analytics
 
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...Boston Institute of Analytics
 
Thiophen Mechanism khhjjjjjjjhhhhhhhhhhh
Thiophen Mechanism khhjjjjjjjhhhhhhhhhhhThiophen Mechanism khhjjjjjjjhhhhhhhhhhh
Thiophen Mechanism khhjjjjjjjhhhhhhhhhhhYasamin16
 
RABBIT: A CLI tool for identifying bots based on their GitHub events.
RABBIT: A CLI tool for identifying bots based on their GitHub events.RABBIT: A CLI tool for identifying bots based on their GitHub events.
RABBIT: A CLI tool for identifying bots based on their GitHub events.natarajan8993
 
Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024Colleen Farrelly
 
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一F sss
 
ASML's Taxonomy Adventure by Daniel Canter
ASML's Taxonomy Adventure by Daniel CanterASML's Taxonomy Adventure by Daniel Canter
ASML's Taxonomy Adventure by Daniel Cantervoginip
 
RadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdfRadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdfgstagge
 
LLMs, LMMs, their Improvement Suggestions and the Path towards AGI
LLMs, LMMs, their Improvement Suggestions and the Path towards AGILLMs, LMMs, their Improvement Suggestions and the Path towards AGI
LLMs, LMMs, their Improvement Suggestions and the Path towards AGIThomas Poetter
 
科罗拉多大学波尔得分校毕业证学位证成绩单-可办理
科罗拉多大学波尔得分校毕业证学位证成绩单-可办理科罗拉多大学波尔得分校毕业证学位证成绩单-可办理
科罗拉多大学波尔得分校毕业证学位证成绩单-可办理e4aez8ss
 
Data Factory in Microsoft Fabric (MsBIP #82)
Data Factory in Microsoft Fabric (MsBIP #82)Data Factory in Microsoft Fabric (MsBIP #82)
Data Factory in Microsoft Fabric (MsBIP #82)Cathrine Wilhelmsen
 
毕业文凭制作#回国入职#diploma#degree澳洲中央昆士兰大学毕业证成绩单pdf电子版制作修改#毕业文凭制作#回国入职#diploma#degree
毕业文凭制作#回国入职#diploma#degree澳洲中央昆士兰大学毕业证成绩单pdf电子版制作修改#毕业文凭制作#回国入职#diploma#degree毕业文凭制作#回国入职#diploma#degree澳洲中央昆士兰大学毕业证成绩单pdf电子版制作修改#毕业文凭制作#回国入职#diploma#degree
毕业文凭制作#回国入职#diploma#degree澳洲中央昆士兰大学毕业证成绩单pdf电子版制作修改#毕业文凭制作#回国入职#diploma#degreeyuu sss
 
detection and classification of knee osteoarthritis.pptx
detection and classification of knee osteoarthritis.pptxdetection and classification of knee osteoarthritis.pptx
detection and classification of knee osteoarthritis.pptxAleenaJamil4
 
原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档
原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档
原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档208367051
 

Último (20)

GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]
 
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝DelhiRS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
 
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
 
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
Call Us ➥97111√47426🤳Call Girls in Aerocity (Delhi NCR)
 
Predictive Analysis for Loan Default Presentation : Data Analysis Project PPT
Predictive Analysis for Loan Default  Presentation : Data Analysis Project PPTPredictive Analysis for Loan Default  Presentation : Data Analysis Project PPT
Predictive Analysis for Loan Default Presentation : Data Analysis Project PPT
 
Data Analysis Project : Targeting the Right Customers, Presentation on Bank M...
Data Analysis Project : Targeting the Right Customers, Presentation on Bank M...Data Analysis Project : Targeting the Right Customers, Presentation on Bank M...
Data Analysis Project : Targeting the Right Customers, Presentation on Bank M...
 
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
 
Thiophen Mechanism khhjjjjjjjhhhhhhhhhhh
Thiophen Mechanism khhjjjjjjjhhhhhhhhhhhThiophen Mechanism khhjjjjjjjhhhhhhhhhhh
Thiophen Mechanism khhjjjjjjjhhhhhhhhhhh
 
RABBIT: A CLI tool for identifying bots based on their GitHub events.
RABBIT: A CLI tool for identifying bots based on their GitHub events.RABBIT: A CLI tool for identifying bots based on their GitHub events.
RABBIT: A CLI tool for identifying bots based on their GitHub events.
 
Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024
 
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
 
ASML's Taxonomy Adventure by Daniel Canter
ASML's Taxonomy Adventure by Daniel CanterASML's Taxonomy Adventure by Daniel Canter
ASML's Taxonomy Adventure by Daniel Canter
 
RadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdfRadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdf
 
LLMs, LMMs, their Improvement Suggestions and the Path towards AGI
LLMs, LMMs, their Improvement Suggestions and the Path towards AGILLMs, LMMs, their Improvement Suggestions and the Path towards AGI
LLMs, LMMs, their Improvement Suggestions and the Path towards AGI
 
Deep Generative Learning for All - The Gen AI Hype (Spring 2024)
Deep Generative Learning for All - The Gen AI Hype (Spring 2024)Deep Generative Learning for All - The Gen AI Hype (Spring 2024)
Deep Generative Learning for All - The Gen AI Hype (Spring 2024)
 
科罗拉多大学波尔得分校毕业证学位证成绩单-可办理
科罗拉多大学波尔得分校毕业证学位证成绩单-可办理科罗拉多大学波尔得分校毕业证学位证成绩单-可办理
科罗拉多大学波尔得分校毕业证学位证成绩单-可办理
 
Data Factory in Microsoft Fabric (MsBIP #82)
Data Factory in Microsoft Fabric (MsBIP #82)Data Factory in Microsoft Fabric (MsBIP #82)
Data Factory in Microsoft Fabric (MsBIP #82)
 
毕业文凭制作#回国入职#diploma#degree澳洲中央昆士兰大学毕业证成绩单pdf电子版制作修改#毕业文凭制作#回国入职#diploma#degree
毕业文凭制作#回国入职#diploma#degree澳洲中央昆士兰大学毕业证成绩单pdf电子版制作修改#毕业文凭制作#回国入职#diploma#degree毕业文凭制作#回国入职#diploma#degree澳洲中央昆士兰大学毕业证成绩单pdf电子版制作修改#毕业文凭制作#回国入职#diploma#degree
毕业文凭制作#回国入职#diploma#degree澳洲中央昆士兰大学毕业证成绩单pdf电子版制作修改#毕业文凭制作#回国入职#diploma#degree
 
detection and classification of knee osteoarthritis.pptx
detection and classification of knee osteoarthritis.pptxdetection and classification of knee osteoarthritis.pptx
detection and classification of knee osteoarthritis.pptx
 
原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档
原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档
原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档
 

Continuous Unsupervised Training of Deep Architectures

  • 1. Continuous Unsupervised Training of Deep Architectures Biometric System Laboratory DISI - University of Bologna http://biolab.csr.unibo.it 6th July 2017 – IIT Genova Davide Maltoni davide.maltoni@unibo.it Vincenzo Lomonaco vincenzo.lomonaco@unibo.it
  • 2. 1Biometric System Laboratory Outline Deep architectures Supervised, Semi supervised and Unsupervised learning Continuous Incremental learning & Forgetting Temporal Coherence for Semi Supervised Learning Core50: Dataset & Benchmark
  • 3. 2Biometric System Laboratory Deep architectures  Hierarchical architectures with many layers Neural Networks (with more than 3 levels) Convolutional Neural Networks (CNN) Hierarchical Temporal Memory (HTM) Multi-stage Hubel-Wiesel Architectures (Ranzato 2007) Alternating layers of feature detectors (Hubel and Wiesel’s simple cells) and local pooling/subsampling of features (Hubel and Wiesel’s complex cells). We focused on CNN and HTM Increasing the number of alternating feature extraction (discrimination) and feature pooling (generalization) layers seems to be the key to tackle invariance Deep architectures (1)
  • 4. 3Biometric System Laboratory State-of-the-art achievements in many fields: vision (object classification, detection, segmentation) speech recognition natural language processing Mainly supervised training (backpropagation like) with huge and fixed datasets: ImageNet - ILSVRC (1.2M labeled images - 1,000 classes) YouTube-8M (7M videos – 4,716 classes) Microsoft Coco (2M instances, 300,000 images – 80 classes) GPU parallelism to speed-up training (months  days) Deep architectures (2)
  • 5. 4Biometric System Laboratory Learning features Learning Features by Watching Objects Move (Pathak et al, 2017) Unsupervised motion-based segmentation as pseudo ground truth Sparse Autoencoder objective function designed to optimize the input reconstruction (promote sparsity and feature grouping) Learning low level feature detectors (V1 like) is quite simple (Olshausen & Field, 1996), but learning high level feature detectors requires huge amount of data (Le 2012, Google) Unsupervised learning reconstruction pooling
  • 6. 5Biometric System Laboratory Semi-supervised Typically a small set of labeled data + large set of unlabeled data. Knowledge of pattern absolute density can help to optimize the decision boundary Self-training a system is first trained with a small amount of labeled data and then used to classify the unlabeled data. The most confident unlabeled points are (iteratively) added to the training set. but unfortunately: patterns whose label can be correctly guessed do not bring much value to improve the current representation really useful patterns (in term of diversity) are not added because of the low self- confidence Semi-supervised training (Xiaojin Zhu, 2007)
  • 7. 6Biometric System Laboratory Adapt an already trained architecture to solve a new problem (in the same domain) Very popular today (avoid time consuming retraining from scratch) Replace (at least) the last layer (top classifier) Reuse of features – Keep them fixed and train only top layers – Tune them together with top layers Is not incremental – Relevant accuracy degradation on the original task Transfer learning
  • 8. 7Biometric System Laboratory Why it is important: Many real-world scenarios are not fixed a priori Training data often available as subsequent batches or from streaming sources Main Strategies: store the past data and retrain the system from scratch (often unrealistic) data comes in sequential batches, used once and then lost (more feasible) Problems: – Catastrophic forgetting (McCloskey & Cohen 1989) – Stability-plasticity dilemma (Mermillod et al. 2013) Continuous - Incremental learning
  • 9. 8Biometric System Laboratory Solutions: – Incremental Classifiers – Camoriano et. al. 2016, Lomonaco & Maltoni 2016* Keep feature extraction fixed and train an incremental classifier (e.g. RLS, SVM) – Early stopping each batch produces only slight changes in the parameters – Self-refreshing generative models to create pseudo-patterns – Elastic Weight Consolidation (EWC) – Kirkpatrick et al. 2016 move only non-critical weights – Learning without Forgetting (LwF) – Li & Hoiem 2016 promote output stability (of the old network on new patterns) Interesting approaches, but tested only on simple (two or few task) problems. Continuous - Incremental learning (2) * Lomonaco and Maltoni, Comparing Incremental Learning Strategies for Convolutional Neural Networks, Workshop on Artificial Neural Networks in Pattern Recognition, 2016.
  • 10. 9Biometric System Laboratory Natural learning Continuous / Lifelong (and possibly online) Partially supervised (or with reinforcement), but mostly unsupervised Multimodal / Multitask human-like learning involves an initial small amount of direct instruction (e.g. parental labeling of objects during childhood) combined with large amounts of subsequence unsupervised experience (e.g. self-interaction with objects) Ultimate interest
  • 11. 10Biometric System Laboratory HTM Hierarchical Temporal Memory (HTM) is a biologically-inspired computational framework proposed by Hawkins and George Dileep George, Jeff Hawkins, “A Hierarchical Bayesian Model of Invariant Pattern Recognition in the Visual Cortex”, IJCNN 2005. Dileep George, Jeff Hawkins, “Towards a Mathematical Theory of Cortical Micro- circuits”, PLoS Comput. Biol. 5(10), 2009. A silicon-valley company (Numenta) created to develop/market HTM technology. HTM review and implementation details in: Davide Maltoni, “Pattern recognition by Hierarchical Temporal Memory”, Tech. Report, DEIS – University of Bologna, April 2011: http://cogprints.org/9187/1/HTM_TR_v1.0.pdf HTM Supervised Refinement (HSR) is a backpropagation-like approach to tune HTM after initial Pre-training: Rehn, E. M. and Maltoni, D. “Incremental learning by message passing in hierarchical temporal memory”. Neural Computation, 26(8):1763–1809, 2014.
  • 12. 11Biometric System Laboratory HTM generations We work on the “first generation” of HTM. In 2011 Numenta focused on a different model (called CLA: Cortical Learning Algorithms) : better suited for dynamic patterns and their temporal relations (compete with Recurrent NN, LSTM). Practical applications of CLA are tools for anomaly detection on data streams (e.g. stock market data, network intrusion detection, etc.). CLA is not efficient enough to work with high-dimensional patterns such as images.
  • 13. 12Biometric System Laboratory HTM Architecture Tree-like network with nlevs (≥ 2) levels: Level 3 (output) 1 node Level 2 (intermediate) 2×2 nodes Level 1 (intermediate) 4×4 nodes Level 0 (input) 16×16 nodes Image 16×16 pixels Example: 4 levels, 16x16 input
  • 14. 13Biometric System Laboratory Node structure Input nodes (level 0) are simple interfaces between pixels and level 1 nodes. Intermediate nodes are the real computation engine of HTM: they store coincidences and temporal groups. A single output node works as a pattern classifiers (e.g., a NN classifiers). 𝛌2 − …𝛌1 − 𝛌 𝑚 − Coincidences C Temporal groups GMatrix 𝐜2 …𝐜1 𝐜 𝑛 𝑐 𝐠2 …𝐠1 𝐠 𝑛 𝑔𝐏𝐂𝐆 𝛌− = 𝛌1 − , 𝛌2 − … 𝛌 𝑚 − 𝐲 𝛌+ 𝛌2 − …𝛌1 − 𝛌 𝑚 − Coincidences C Prior class prob.Matrix 𝐜2 …𝐜1 𝐜 𝑛 𝑐 [𝑃(𝑤1), 𝑃(𝑤2) … 𝑃(𝑤𝑛 𝑤 )]𝐏𝐂𝐖 𝛌− = 𝛌1 − , 𝛌2 − … 𝛌 𝑚 − 𝐲 [𝑃(𝑤1|𝑒), 𝑃(𝑤2|𝑒) … 𝑃(𝑤𝑛 𝑤 |𝑒)] Intermediate node Output node
  • 15. 14Biometric System Laboratory Coincidences Each coincidence is a sort of feature extractor that spans a portion of the image corresponding to the node receptive field (i.e., small at low levels and large at high levels). Coincidences are used to perform a spatial analysis of input patterns and to find out spatial similarities. Example of level 1 coincidences
  • 16. 15Biometric System Laboratory Temporal groups A temporal group is a subset of coincidences, that could be spatially quite different each from the other, but that are likely to be originated from simple variations of the same pattern. HTM exploits temporal smoothness to create temporal groups: patterns presented to the network very close in time, are likely to be variants of the same pattern that is smoothly moving throughout the network field of view. Example of level 1 groups 𝐠1 𝐠2 𝐠3 𝐠4 𝐠5 𝐠6 𝐠7 𝐠8 𝐠9 𝐠10 𝐠11 𝐠12 𝐠13 𝐠14
  • 17. 16Biometric System Laboratory HTM vs CNN Key points: Bayesian probabilistic formulation: Bayesian Network + Bayesian Belief propagation equations. Parameters values constrained by probability laws. Top down and bottom-up information flow: feedback messages from higher levels carry contextual information to bias the behavior of lower levels. By fusing bottom up and top down messages each HTM node reaches an internal state (called node belief and corresponding to Bayes posterior) which is an optimal probabilistic explanation of the external stimuli. Pooling is a key component in HTM (some CNN work well even without pooling layers). Unsupervised Pre-training is very effective in HTM. Compares favorably on small-scale problems, difficult to scale-up. Designing large HTM is not easy (architectural choices, pre-training, code optimization, numerical problems). 16x16  32x32  64x64  128x128  …
  • 18. 17Biometric System Laboratory From supervised to Semi-Supervised Tuning Exploiting time (again) as supervisor. Biological plausibility: … hypothesis under which invariance is learned from temporal continuity of object features during natural visual experience without external supervision (Li & DiCarlo 2008) SST idea: pass back the current output vector as desired vector (i.e., label) for the next pattern D. Maltoni and V. Lomonaco, Semi-supervised Tuning from Temporal Coherence, International Conference on Pattern Recognition, 2016. ArXiv 1511.03163
  • 19. 18Biometric System Laboratory SST (1) Temporal coherent patterns: 𝒗 𝑡 , 𝑡 = 1 … 𝑚 A classifier 𝑵 maps an input pattern to an output vector: 𝒗 𝑡 → 𝑵 𝒗 𝑡 Loss function to be minimized (e.g. backprop): 1 2 𝑵 𝒗 𝑡 − 𝒅(𝒗 𝑡 ) 2 1. Supervised Tuning (SupT): 𝒅 𝒗 𝑡 = 𝚫 𝑤 = 0, … , 1, … 0 w is the pattern class 2. Supervised Tuning with Regularization (SupTR): 𝒅 𝒗 𝑡 = 𝜆 ∙ 𝚫 𝑤 + 1 − 𝜆 ∙ 𝑵 𝒗 𝑡−1 position w
  • 20. 19Biometric System Laboratory SST (2) 3. Semi Supervised Tuning – Basic (SST-B): 𝒅 𝒗 𝑡 = 𝑵 𝒗 𝑡−1 4. Semi Supervised Tuning – Advanced (SST-A): 𝒇 𝒗 𝑡 = 𝑵 𝒗 𝑡−1 𝑡 = 2 𝒇 𝒗 𝑡−1 + 𝑵 𝒗 𝑡−1 2 𝑡 > 2 𝒅 𝒗 𝑡 = 𝒇 𝒗 𝑡 𝑖𝑓 max 𝑖 𝒇𝑖 𝒗 𝑡 > 𝑠𝑐 𝑵 𝒗 𝑡 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
  • 21. 20Biometric System Laboratory SST Experiments We generated temporal coherent sequences from NORB and COIL100 datasets by randomly walking the variation space (pose, lighting). 1 sequence for initial (supervised training) + 9 sequences for incremental tuning Test set contains frames with a given mindist from all training patterns
  • 22. 21Biometric System Laboratory SST Experiments (2) HTM on NORB CNN on NORB
  • 23. 22Biometric System Laboratory SST Experiments (3) making the problem harder by reducing initial accuracy … Comparison with self-training Self-training
  • 24. 23Biometric System Laboratory Ongoing work on SST SST approach very effective with HTM SST-A accuracy is sometime very close to supervised scenario. Why does HTM work better than CNN with SST? HTM seems to be more robust w.r.t. forgetting, Why? In HTM we keeps coincidence fixed. Try with CNN - NiN (Network in Network) models where we move only feature pooling weights. Investigate other architectural differences. New experiments on natural datasets and continuous learning scenarios. Core50 iCubWorld-Transf
  • 25. 24Biometric System Laboratory CORe50 Dataset, Benchmark, code and additional information freely available at: https://vlomonaco.github.io/core50
  • 26. 25Biometric System Laboratory CORe50 (2) Motivation Temporal Coherence has already shown to be a good surrogate supervised signal (also in complex sequence learning problems) Still, it’s very difficult to find simple videos of objects smoothly moving in front of the camera. Moreover, in order to assess Continuous Learning scenarios we need the presence of multiple (temporal coherent and unconstrained) views of the same objects taken in different sessions (varying background, lighting, pose, occlusions, etc.) Sometimes it is possible to generate a number of exploration sequences that turn a native static benchmarks into continuous learning tasks, but the result is often unnatural (Maltoni & Lomonaco 2016)
  • 28. 27Biometric System Laboratory 50 Objects of 10 Classes
  • 29. 28Biometric System Laboratory 11 Sessions One frame of the same object (#41) throughout the 11 acquisition sessions. Three of the eleven sessions (#3, #7 and #10) have been selected for test and the remaining 8 sessions are used for training.
  • 30. 29Biometric System Laboratory Acquisition Setting For each session and for each object, a 15 seconds video (at 20 fps) has been recorded with a Kinect 2.0 sensor delivering 300 RGB-D frames. The acquisition interface identifies a central region where the object should be this allows to crop the frame size from 1024×575 to 350×350.
  • 31. 30Biometric System Laboratory Acquisition setting (2) The 350×350 frames are then automatically cropped to 128×128 based on a fully automated tracker. Example of 1 second recording (at 20 fps) of object #26 in session #4 (outdoor). Note the smooth movement, pose change and partial occlusion.
  • 32. 31Biometric System Laboratory Continuous Learning Scenarios New Instances (NI) New training patterns of the same classes become available in subsequent batches with new poses and conditions (illumination, background, occlusion, etc.). New Classes (NC) New training patterns belonging to different classes become available in subsequent batches. New Instances and Classes (NIC) New training patterns belonging both to known and new classes become available in subsequent training batches.
  • 33. 32Biometric System Laboratory Continuous Learning Strategies Cumulative (non-continuous) We re-train the entire model from scratch as soon as a new batch of data is available. Naïve We simply continue back-propagation with early-stopping and low learning rate. Copy Weights with Re-init (CWR) Simple baseline to add new classes easily disentangling the weights effecting each class.
  • 34. 33Biometric System Laboratory Copy Weights with Re-init (CWR) In CWR we skip layers fc6 and fc7 and directly connect pool5 to a final layer fc8 (followed by softmax) while maintaining the weights up to Pool5 fixed. This allows isolating the subsets of weights that each class uses. During the training two sets of weights are maintained by the model for the pool5→fc8 connections: cw are the consolidated weights used for inference and tw the temporary weights used for training. After the training of each batch temporary weights tw are copied in cw. twcwfixed
  • 35. 34Biometric System Laboratory Copy Weights with Re-init (CWR) In CWR we skip layers fc6 and fc7 and directly connect pool5 to a final layer fc8 (followed by softmax) while maintaining the weights up to Pool5 fixed. This allows isolating the subsets of weights that each class uses. During the training two sets of weights are maintained by the model for the pool5→fc8 connections: cw are the consolidated weights used for inference and tw the temporary weights used for training. After the training of each batch temporary weights tw are copied in cw. twcwfixed
  • 36. 35Biometric System Laboratory Copy Weights with Re-init (CWR) In CWR we skip layers fc6 and fc7 and directly connect pool5 to a final layer fc8 (followed by softmax) while maintaining the weights up to Pool5 fixed. This allows isolating the subsets of weights that each class uses. During the training two sets of weights are maintained by the model for the pool5→fc8 connections: cw are the consolidated weights used for inference and tw the temporary weights used for training. After the training of each batch temporary weights tw are copied in cw. twcwfixed
  • 37. 36Biometric System Laboratory Copy Weights with Re-init (CWR) In CWR we skip layers fc6 and fc7 and directly connect pool5 to a final layer fc8 (followed by softmax) while maintaining the weights up to Pool5 fixed. This allows isolating the subsets of weights that each class uses. During the training two sets of weights are maintained by the model for the pool5→fc8 connections: cw are the consolidated weights used for inference and tw the temporary weights used for training. After the training of each batch temporary weights tw are copied in cw. twcwfixed
  • 38. 37Biometric System Laboratory Copy Weights with Re-init (CWR) In CWR we skip layers fc6 and fc7 and directly connect pool5 to a final layer fc8 (followed by softmax) while maintaining the weights up to Pool5 fixed. This allows isolating the subsets of weights that each class uses. During the training two sets of weights are maintained by the model for the pool5→fc8 connections: cw are the consolidated weights used for inference and tw the temporary weights used for training. After the training of each batch temporary weights tw are copied in cw. twcwfixed
  • 39. 38Biometric System Laboratory Copy Weights with Re-init (CWR) In CWR we skip layers fc6 and fc7 and directly connect pool5 to a final layer fc8 (followed by softmax) while maintaining the weights up to Pool5 fixed. This allows isolating the subsets of weights that each class uses. During the training two sets of weights are maintained by the model for the pool5→fc8 connections: cw are the consolidated weights used for inference and tw the temporary weights used for training. After the training of each batch temporary weights tw are copied in cw. twcwfixed
  • 40. 39Biometric System Laboratory New Instances (NI) In this scenario the training batches coincides with the 8 sessions available in the training set. For each of the 50 objects, training a model on the first session ad tuning it 7 times (on the remaining 7 sessions).
  • 41. 40Biometric System Laboratory New Classes (NC) Each batch contains the whole training sequences (8) of a small group of classes, and therefore no memory refresh is possible across batches. In the first batch we include 10 classes, while the remaining 8 batches contain 5 classes each.
  • 42. 41Biometric System Laboratory New Instances and Classes (NIC) The first batch includes 10 classes, and the subsequent batches 5 classes each. However, only one training sequence per class is here included in a batch, thus resulting in a double partitioning scheme (i.e., classes and sequences). The total number of batches is 79.
  • 43. 42Biometric System Laboratory Conclusions Continuous/Lifelong Learning enables a boarder range of real-word applications and counts a number of different advantages Biologically grounded and memory/computational efficient Enables adaptation and knowledge reuse/refining Ideal fit for unsupervised streaming perception data (Visual, Audio, etc..) Focus on all the three main scenarios NI, NC and NIC not just NC Recent literature on Continuous/Lifelong learning focuses on simple sequences of different tasks (maximum 3-4). Temporal coherence is one of the key towards unsupervised continuous learning Sequence learning, Semi-Supervised Tuning (SST), etc…
  • 44. 43Biometric System Laboratory Future Plans Continuing out work in SST and CORe50 HTM vs CNN comparison Further tests on CORe50 and iCubWorld-Transf Implementing and evaluating LwF and EWC on Core50 LwF and EWC have been only tested in a NC-like scenario. It is not trivial to apply them in the NI and NIC scenarios.
  • 45. Thank you for listening! Biometric System Laboratory DISI - University of Bologna http://biolab.csr.unibo.it 6th July 2017 – IIT Genova Davide Maltoni davide.maltoni@unibo.it Vincenzo Lomonaco vincenzo.lomonaco@unibo.it Continuous Unsupervised Training of Deep Architectures
  • 46. 45Biometric System Laboratory CNN incremental (supervised) training (2) incremental Strategy day1 day2 day3 Lenet7 36,00% 37,55% 35,84% AlexNet + RLS 58,20% 66,39% 74,10% AlexNet + SVM 56,32% 66,19% 72,07% FRCNN + Svm 51,63% 59,63% 70,00% AlexNet + finetuning (last lvl) 54,01% 60,00% 65,38% AlexNet + finetuning 70,70% 77,14% 76,85% IcubWorld28 Img_size: 128x128 Inc_batches: 3 Num_classes: 5 (x4 obj) Time
  • 47. Time 46Biometric System Laboratory CNN incremental (supervised) training (3) Big-brother (setB) Img_size: 70x70 Inc_batches: 56 Num_classes: 7