Complexity and Computation in Nature: How can we test for Artificial Life?Hector Zenil
Invited Guest Lecture on Complexity and Natural Computing for the ShangAI Lectures on Natural and Artificial Intelligence. Broadcasted simultaneously to about 20 universities: UC3 Madrid; Zurich; Pisa, Humboldt, Berlin; Plymouth & Salford, UK; China and Japan, November 7, 2013. Video available at https://cast.switch.ch/vod/clips/2m1xxl21ej/
Complexity and Computation in Nature: How can we test for Artificial Life?Hector Zenil
Invited Guest Lecture on Complexity and Natural Computing for the ShangAI Lectures on Natural and Artificial Intelligence. Broadcasted simultaneously to about 20 universities: UC3 Madrid; Zurich; Pisa, Humboldt, Berlin; Plymouth & Salford, UK; China and Japan, November 7, 2013. Video available at https://cast.switch.ch/vod/clips/2m1xxl21ej/
Large Convolutional Network models have
recently demonstrated impressive classification
performance on the ImageNet benchmark
(Krizhevsky et al., 2012). However
there is no clear understanding of why they
perform so well, or how they might be improved.
In this paper we address both issues.
We introduce a novel visualization technique
that gives insight into the function of intermediate
feature layers and the operation of
the classifier. Used in a diagnostic role, these
visualizations allow us to find model architectures
that outperform Krizhevsky et al. on
the ImageNet classification benchmark. We
also perform an ablation study to discover
the performance contribution from different
model layers. We show our ImageNet model
generalizes well to other datasets: when the
softmax classifier is retrained, it convincingly
beats the current state-of-the-art results on
Caltech-101 and Caltech-256 datasets
Cognition, Information and Subjective ComputationHector Zenil
One of the most important contending theories deeply connects consciousness to information theory. We keep connecting mind properties to computation. Turing did it with human intelligence and computation. John Searle (unintended I will claim) connected understanding (and consciousness) to program complexity (and soft AI). And more recently, Guilio Tononi formally connected internal experience and consciousness to computation and information. Therefore, can understanding computation shed light on intelligence and consciousness? I claim it does. So what is computation?. I aim at finding a grading (such as Tononi's phi) metric of computation, weakly observer dependent (following some ideas of Searle) and with considerations to resources complexity to give it sense to the Turing test (as Scott Aaronson would agree with).
I think this could be useful for those who works in the field of Coputational Intelligence. Give your valuable reviews so that I can progree in my research
Offline Character Recognition Using Monte Carlo Method and Neural Networkijaia
Human Machine interface are constantly gaining improvements because of increasing development of
computer tools. Handwritten Character Recognition do have various significant applications like form
scanning, verification, validation, or checks reading. Because of the importance of these applications
passionate research in the field of Off-Line handwritten character recognition is going on. The challenge in
recognising the handwritings lies in the nature of humans, having unique styles in terms of font, contours,
etc. This paper presents a novice approach to identify the offline characters; we call it as character divider
approach which can be used after pre-processing stage. We devise an innovative approach for feature
extraction known as vector contour. We also discuss the pros and cons including limitations, of our
approach
Complex systems are characterized by constituents -- from neurons in the brain to individuals in a social network -- which exhibit special structural organization and nonlinear dynamics. As a consequence, a complex system cannot be understood by studying its units separately because their interactions lead to unexpected emerging phenomena, from collective behavior to phase transitions.
Recently, we have discovered that a new level of complexity characterizes a variety of natural and artificial systems, where units interact, simultaneously, in distinct ways. For instance, this is the case of multimodal transportation systems (e.g., metro, bus and train networks) or of biological molecules, whose interactions might be of different type (e.g. physical, chemical, genetic) or functionality (e.g., regulatory, inhibitory, etc.). The unprecedented newfound wealth of multivariate data allows to categorize system's interdependency by defining distinct "layers", each one encoding a different network representation of the system. The result is a multilayer network model.
Analyzing data from different domains -- including molecular biology, neuroscience, urban transport, telecommunications -- we will show that neglecting or disregarding multivariate information might lead to poor results. Conversely, multilayer models provide a suitable framework for complex data analytics, allowing to quantify the resilience of a system to perturbations (e.g., localized failures or targeted attacks), improving forecasting of spreading processes and accuracy in classification problems.
Performance Evaluation of Object Tracking Technique Based on Position VectorsCSCJournals
In this paper, a novel algorithm for moving object tracking based on position vectors has proposed. The position vector of an object in first frame of a video has been extracted based on selection of region of interest. Based on position vector in first frame object direction has shown in nine different directions. We extract nine position vectors for nine different directions. With these position vectors next frame is cropped into nine blocks. We exploit block matching of the first frame with nine blocks of the next frame in a simple feature space by Descrete wavelet transform and dual tree complex wavelet transform. The matched block is considered as tracked object and its position vector is a reference location for the next successive frame. We describe performance evaluation and algorithm in detail to perform simulation experiments of object tracking using different feature vectors which verifies the tracking algorithm efficiency.
Piotr Mirowski - Review Autoencoders (Deep Learning) - CIUUK14Daniel Lewis
Piotr Mirowski (of Microsoft Bing London) presented Review of Auto-Encoders to the Computational Intelligence Unconference 2014, with our Deep Learning stream. These are his slides. Original link here: https://piotrmirowski.files.wordpress.com/2014/08/piotrmirowski_ciunconf_2014_reviewautoencoders.pptx
He also has Matlab-based tutorial on auto-encoders available here:
https://github.com/piotrmirowski/Tutorial_AutoEncoders/
Neural Network Fundamentals being explained using most basic form of mathematics. The slides slowly dives into solid geometry to explain few fundamental concepts on which gradient descent and steepest descent method works.
Continuum Modeling and Control of Large Nonuniform NetworksYang Zhang
Presented at The 49th Annual Allerton Conference on Communication, Control, and Computing, 2011
Abstract—Recent research has shown that some Markov chains modeling networks converge to continuum limits, which are solutions of partial differential equations (PDEs), as the number of the network nodes approaches infinity. Hence we can approximate such large networks by PDEs. However, the previous results were limited to uniform immobile networks with a fixed transmission rule. In this paper we first extend the analysis to uniform networks with more general transmission rules. Then through location transformations we derive the continuum limits of nonuniform and possibly mobile networks. Finally, by comparing the continuum limits of corresponding nonuniform and uniform networks, we develop a method to control the transmissions in nonuniform and mobile networks so that the continuum limit is invariant under node locations, and hence mobility. This enables nonuniform and mobile networks to maintain stable global characteristics in the presence of varying node locations.
Large Convolutional Network models have
recently demonstrated impressive classification
performance on the ImageNet benchmark
(Krizhevsky et al., 2012). However
there is no clear understanding of why they
perform so well, or how they might be improved.
In this paper we address both issues.
We introduce a novel visualization technique
that gives insight into the function of intermediate
feature layers and the operation of
the classifier. Used in a diagnostic role, these
visualizations allow us to find model architectures
that outperform Krizhevsky et al. on
the ImageNet classification benchmark. We
also perform an ablation study to discover
the performance contribution from different
model layers. We show our ImageNet model
generalizes well to other datasets: when the
softmax classifier is retrained, it convincingly
beats the current state-of-the-art results on
Caltech-101 and Caltech-256 datasets
Cognition, Information and Subjective ComputationHector Zenil
One of the most important contending theories deeply connects consciousness to information theory. We keep connecting mind properties to computation. Turing did it with human intelligence and computation. John Searle (unintended I will claim) connected understanding (and consciousness) to program complexity (and soft AI). And more recently, Guilio Tononi formally connected internal experience and consciousness to computation and information. Therefore, can understanding computation shed light on intelligence and consciousness? I claim it does. So what is computation?. I aim at finding a grading (such as Tononi's phi) metric of computation, weakly observer dependent (following some ideas of Searle) and with considerations to resources complexity to give it sense to the Turing test (as Scott Aaronson would agree with).
I think this could be useful for those who works in the field of Coputational Intelligence. Give your valuable reviews so that I can progree in my research
Offline Character Recognition Using Monte Carlo Method and Neural Networkijaia
Human Machine interface are constantly gaining improvements because of increasing development of
computer tools. Handwritten Character Recognition do have various significant applications like form
scanning, verification, validation, or checks reading. Because of the importance of these applications
passionate research in the field of Off-Line handwritten character recognition is going on. The challenge in
recognising the handwritings lies in the nature of humans, having unique styles in terms of font, contours,
etc. This paper presents a novice approach to identify the offline characters; we call it as character divider
approach which can be used after pre-processing stage. We devise an innovative approach for feature
extraction known as vector contour. We also discuss the pros and cons including limitations, of our
approach
Complex systems are characterized by constituents -- from neurons in the brain to individuals in a social network -- which exhibit special structural organization and nonlinear dynamics. As a consequence, a complex system cannot be understood by studying its units separately because their interactions lead to unexpected emerging phenomena, from collective behavior to phase transitions.
Recently, we have discovered that a new level of complexity characterizes a variety of natural and artificial systems, where units interact, simultaneously, in distinct ways. For instance, this is the case of multimodal transportation systems (e.g., metro, bus and train networks) or of biological molecules, whose interactions might be of different type (e.g. physical, chemical, genetic) or functionality (e.g., regulatory, inhibitory, etc.). The unprecedented newfound wealth of multivariate data allows to categorize system's interdependency by defining distinct "layers", each one encoding a different network representation of the system. The result is a multilayer network model.
Analyzing data from different domains -- including molecular biology, neuroscience, urban transport, telecommunications -- we will show that neglecting or disregarding multivariate information might lead to poor results. Conversely, multilayer models provide a suitable framework for complex data analytics, allowing to quantify the resilience of a system to perturbations (e.g., localized failures or targeted attacks), improving forecasting of spreading processes and accuracy in classification problems.
Performance Evaluation of Object Tracking Technique Based on Position VectorsCSCJournals
In this paper, a novel algorithm for moving object tracking based on position vectors has proposed. The position vector of an object in first frame of a video has been extracted based on selection of region of interest. Based on position vector in first frame object direction has shown in nine different directions. We extract nine position vectors for nine different directions. With these position vectors next frame is cropped into nine blocks. We exploit block matching of the first frame with nine blocks of the next frame in a simple feature space by Descrete wavelet transform and dual tree complex wavelet transform. The matched block is considered as tracked object and its position vector is a reference location for the next successive frame. We describe performance evaluation and algorithm in detail to perform simulation experiments of object tracking using different feature vectors which verifies the tracking algorithm efficiency.
Piotr Mirowski - Review Autoencoders (Deep Learning) - CIUUK14Daniel Lewis
Piotr Mirowski (of Microsoft Bing London) presented Review of Auto-Encoders to the Computational Intelligence Unconference 2014, with our Deep Learning stream. These are his slides. Original link here: https://piotrmirowski.files.wordpress.com/2014/08/piotrmirowski_ciunconf_2014_reviewautoencoders.pptx
He also has Matlab-based tutorial on auto-encoders available here:
https://github.com/piotrmirowski/Tutorial_AutoEncoders/
Neural Network Fundamentals being explained using most basic form of mathematics. The slides slowly dives into solid geometry to explain few fundamental concepts on which gradient descent and steepest descent method works.
Continuum Modeling and Control of Large Nonuniform NetworksYang Zhang
Presented at The 49th Annual Allerton Conference on Communication, Control, and Computing, 2011
Abstract—Recent research has shown that some Markov chains modeling networks converge to continuum limits, which are solutions of partial differential equations (PDEs), as the number of the network nodes approaches infinity. Hence we can approximate such large networks by PDEs. However, the previous results were limited to uniform immobile networks with a fixed transmission rule. In this paper we first extend the analysis to uniform networks with more general transmission rules. Then through location transformations we derive the continuum limits of nonuniform and possibly mobile networks. Finally, by comparing the continuum limits of corresponding nonuniform and uniform networks, we develop a method to control the transmissions in nonuniform and mobile networks so that the continuum limit is invariant under node locations, and hence mobility. This enables nonuniform and mobile networks to maintain stable global characteristics in the presence of varying node locations.
In this article we consider macrocanonical models for texture synthesis. In these models samples are generated given an input texture image and a set of features which should be matched in expectation. It is known that if the images are quantized, macrocanonical models are given by Gibbs measures, using the maximum entropy principle. We study conditions under which this result extends to real-valued images. If these conditions hold, finding a macrocanonical model amounts to minimizing a convex function and sampling from an associated Gibbs measure. We analyze an algorithm which alternates between sampling and minimizing. We present experiments with neural network features and study the drawbacks and advantages of using this sampling scheme.
Representing Simplicial Complexes with MangrovesDavid Canino
These slides have been presented at the 22nd International Meshing Roundtable, Orlando, FL, USA. They describe our GPL software, the Mangrove TDS Library: http://mangrovetds.sourceforge.net. It is a C++ tool for the fast prototyping of topological data structures, representing dynamically simplicial and cell complexes.
안녕하세요 딥논읽-DNR입니다!
오늘 소개드릴 논문은 'YOLOS' 라는 논문입니다. YOLOS에 대해 간략하게 먼저 설명을 드리면 오직 Transformer만을 이용하여 2D object Detection을 수행한 모델이라고 이해해 주시면 됩니다. 구조는 오직 Transformer의 Encoder만을 사용하여 Object detection을 수행하였는대요, 데이터셋을 균등하게 학습시켜도 오브젝트마다의 AP가 차이가 심했던 다른 CNN기반의 Object detector와 다르게, 해당 모델은 모든 카테고리에 대해서 AP가 꽤나 균등하게 나오는것도 중요한 특징중 하나 입니다.
오늘 논문 리뷰를 이미지 처리팀 김병현님이 자세한 리뷰를 도와주셨습니다! 오늘도 많은 관심 미리 감사드립니다!
A full description of the molecular autoencoder for automated exploration of chemical compound space using neural nets and machine learning architectures, developed by the Aspuru-Guzik group at Harvard. Talk given to Prof. Peter W. Chung's research group at the University of Maryland, College Park, August 2017.
A Simple Introduction to Neural Information RetrievalBhaskar Mitra
Neural Information Retrieval (or neural IR) is the application of shallow or deep neural networks to IR tasks. In this lecture, we will cover some of the fundamentals of neural representation learning for text retrieval. We will also discuss some of the recent advances in the applications of deep neural architectures to retrieval tasks.
(These slides were presented at a lecture as part of the Information Retrieval and Data Mining course taught at UCL.)
Dynamic stiffness and eigenvalues of nonlocal nano beams - new methods for dynamic analysis of nano-scale structures. This lecture gives a review and proposed new techniques.
Unit 8 - Information and Communication Technology (Paper I).pdfThiyagu K
This slides describes the basic concepts of ICT, basics of Email, Emerging Technology and Digital Initiatives in Education. This presentations aligns with the UGC Paper I syllabus.
Read| The latest issue of The Challenger is here! We are thrilled to announce that our school paper has qualified for the NATIONAL SCHOOLS PRESS CONFERENCE (NSPC) 2024. Thank you for your unwavering support and trust. Dive into the stories that made us stand out!
Francesca Gottschalk - How can education support child empowerment.pptxEduSkills OECD
Francesca Gottschalk from the OECD’s Centre for Educational Research and Innovation presents at the Ask an Expert Webinar: How can education support child empowerment?
Operation “Blue Star” is the only event in the history of Independent India where the state went into war with its own people. Even after about 40 years it is not clear if it was culmination of states anger over people of the region, a political game of power or start of dictatorial chapter in the democratic setup.
The people of Punjab felt alienated from main stream due to denial of their just demands during a long democratic struggle since independence. As it happen all over the word, it led to militant struggle with great loss of lives of military, police and civilian personnel. Killing of Indira Gandhi and massacre of innocent Sikhs in Delhi and other India cities was also associated with this movement.
Safalta Digital marketing institute in Noida, provide complete applications that encompass a huge range of virtual advertising and marketing additives, which includes search engine optimization, virtual communication advertising, pay-per-click on marketing, content material advertising, internet analytics, and greater. These university courses are designed for students who possess a comprehensive understanding of virtual marketing strategies and attributes.Safalta Digital Marketing Institute in Noida is a first choice for young individuals or students who are looking to start their careers in the field of digital advertising. The institute gives specialized courses designed and certification.
for beginners, providing thorough training in areas such as SEO, digital communication marketing, and PPC training in Noida. After finishing the program, students receive the certifications recognised by top different universitie, setting a strong foundation for a successful career in digital marketing.
Executive Directors Chat Leveraging AI for Diversity, Equity, and InclusionTechSoup
Let’s explore the intersection of technology and equity in the final session of our DEI series. Discover how AI tools, like ChatGPT, can be used to support and enhance your nonprofit's DEI initiatives. Participants will gain insights into practical AI applications and get tips for leveraging technology to advance their DEI goals.
A Strategic Approach: GenAI in EducationPeter Windle
Artificial Intelligence (AI) technologies such as Generative AI, Image Generators and Large Language Models have had a dramatic impact on teaching, learning and assessment over the past 18 months. The most immediate threat AI posed was to Academic Integrity with Higher Education Institutes (HEIs) focusing their efforts on combating the use of GenAI in assessment. Guidelines were developed for staff and students, policies put in place too. Innovative educators have forged paths in the use of Generative AI for teaching, learning and assessments leading to pockets of transformation springing up across HEIs, often with little or no top-down guidance, support or direction.
This Gasta posits a strategic approach to integrating AI into HEIs to prepare staff, students and the curriculum for an evolving world and workplace. We will highlight the advantages of working with these technologies beyond the realm of teaching, learning and assessment by considering prompt engineering skills, industry impact, curriculum changes, and the need for staff upskilling. In contrast, not engaging strategically with Generative AI poses risks, including falling behind peers, missed opportunities and failing to ensure our graduates remain employable. The rapid evolution of AI technologies necessitates a proactive and strategic approach if we are to remain relevant.
4. !4
Transform Learning(TL)
Transform Original Data Sparse Representation
k × m m × n k × n
The transform learning1 problem can be expressed as
min
T∈!k×m
,Z∈!k×n
||TX − Z ||F
2
+λ(ε ||T ||F
2
−logdetT) + µ || Z ||0
1. S. Ravishankar and Y. Bresler, “Learning Sparsifying Transforms”, IEEE Transactions on Signal Processing, (2013).
(T) (X) (Z)
TX = Z
5. !5
Transform Learning(TL)
min
T∈!k×m
,Z∈!k×n
||TX − Z ||F
2
+λ(ε ||T ||F
2
−logdetT) + µ || Z ||0
T ← min
T
||TX − Z ||F
2
+λ(ε ||T ||F
2
−logdetT )
Z ← min
Z
||TX − Z ||F
2
+µ || Z ||0
Closed form
solution exists
6. !6
Transform Learning(TL)
min
T∈!k×m
,Z∈!k×n
||TX − Z ||F
2
+λ(ε ||T ||F
2
−logdetT) + µ || Z ||0
T ← min
T
||TX − Z ||F
2
+λ(ε ||T ||F
2
−logdetT )
Z ← min
Z
||TX − Z ||F
2
+µ || Z ||0
Closed form
solution exists
Idea : Use transforms for solving Machine Learning Problems??
9. Supervised Transform Learning
Label Consistent TL (LCTL):
!9
! min
T ,Z,M
||TX − Z ||F
2
+λ(ε ||T ||F
2
−logdetT )+ µ || Z ||1 +η ||Q − MZ ||F
2
supervision term
Learn mapping between true labels Q and the coefficients Z.
Mapping M can be linear or non-linear.
X Z
Q
T
M
10. Kernel Transform Learning
Kernel TL:
• Dense transform —> non-linear transformation of noisy
data into higher dimensional feature space .X ϕ
!10
: fixed basis
: transform as sparse
combination (B) of basis from
Φ
ΦB
Φ
ϕ :!N
→ F
TX = Z
K(X, X) = ϕ(X)T
ϕ(X)
BK(X,X) = Z
Bϕ(X)T
Transform
!"# $#
ϕ(X)
Data
! = Z
11. Supervised Kernel Transform Learning
Transform Learning:
Kernel TL:
Kernel LCTL:
!11
min
B,Z,M
|| BK(X, X)− Z ||F
2
+λ(ε ||T ||F
2
−logdetT )+ µ || Z ||1 +η ||Q − MZ ||F
2
min
B,Z
|| BK(X, X)− Z ||F
2
+λ(ε || B ||F
2
−logdet B)+ µ || Z ||0
supervision term
Kernel transform term
Kernel transform term
min
T ,Z
||TX − Z ||F
2
+λ(ε ||T ||F
2
−logdetT )+ µ || Z ||1
12. Supervised Transform Learning: Results
• Classification Results on YaleB(38 persons), AR faces(100 persons)
• Kernel: polynomial order 3
• Parameter values obtained on CIFAR-10 validation dataset
• Benchmark comparisons:
• Discriminative Baysian Dictionary Learning(DBDL)2
• Multimodal Task Driven Dictionary Learning(MTDL)3
• Discriminative Analysis Dictionary Learning(DADL)4
• Sparse Embedded Dictionary Learning(SEDL)5
• Non-Linear Dictionary Learning(NDL)6
!12
2. N. Akhtar, F. Shafait and A. Mian, "Discriminative Bayesian Dictionary Learning for Classification," IEEE Transactions on Pattern
Analysis and Machine Intelligence, 2016.
3. S. Bahrampour, N. M. Nasrabadi, A. Ray and W. K. Jenkins, "Multimodal Task-Driven Dictionary Learning for Image Classification,"
IEEE Transactions on Image Processing, 2016.
4. J. Guo, Y. Guo, X. Kong, M. Zhang and R. He, “Discriminative Analysis Dictionary Learning”, AAAI Conference on Artificial
Intelligence, 2016.
5. Y. Chen and J. Su, “Sparse embedded dictionary learning on face recognition”, Pattern Recognition, 2017.
6. J. Hu and Y.-P. Tan, “Nonlinear dictionary learning with application to image classification”, Pattern Recognition (in Press).
15. Deep Transform Learning(DTL)
!15
Basic Idea : Repeat transforms to form a deeper architecture
To learn N-levels of transform, the model is
TN
ϕ...(T2
ϕ(T1
X )) = Z
All layers are
learned jointly
16. Jointly Learned Deep Transform Learning
!16
min
T1,T2 ,Z1,Z
||T2Z1 − Z ||F
2
+λ (µ ||Ti ||F
2
−logdetTi )
i=1
2
∑ + µ ||T1X −φ−1
Z1( )||F
2
● Formulation for two layer network
● Variable splitting
● All coefficients and transforms are learned in one loop
min
T1,T2 ,Z
||T2 (φ(T1X))− Z ||F
2
+λ (µ ||Ti ||F
2
−logdetTi )
i=1
2
∑
Z1 = φ(T1X)
T2Z1 = Z
17. Jointly Learned Deep Transform Learning
● S1 :
● S2 :
● S3 :
● S4 :
!17
min
T1,T2 ,Z1,Z
||T2Z1 − Z ||F
2
+λ (µ ||Ti ||F
2
−logdetTi )
i=1
2
∑ + µ ||T1X −φ−1
Z1( )||F
2
min
T2
||T2Z1 − Z ||F
2
+λ(||T2 ||F
2
−logdetT2 )
min
T1
µ ||T1X −φ−1
(Z1)||F
2
+λ(||T1 ||F
2
−logdetT1)
min
Z
||T2Z1 − Z ||F
2
⇒ T2Z1 = Z
min
Z1
||T2Z1 − Z ||F
2
+µ ||φ(T1X)− Z1 ||F
2
18. Classification Results: Joint DTL
!18
Classification Accuracy with SVM
Method YALEB AR Faces
CSSAE7 85.21 82.22
CSDBN8 84.97 82.11
DDL9 92.66 93.35
Proposed 1-layer 95.11 94.98
Proposed 2-layers 97.41 95.87
Proposed 3-layers 97.67 96.80
Proposed 4-layers 96.36 96.24
7. A. Sankaran, M. Vatsa, R. Singh, and A. Majumdar, “Group sparse autoencoder,” Image and Vision Computing, 2017.
8. A. Sankaran, G. Goswami, M. Vatsa, R. Singh, and A. Majumdar, “Class sparsity signature based restricted boltzmann
machine,” Pattern Recognition, 2017.
9. V. Singal and A. Majumdar, "Majorization Minimization Technique for Optimally Solving Deep Dictionary Learning",
Neural Processing Letters, doi:10.1007/s11063-017-9603-9, 2017.
19. Clustering Results
!19
K-Means: YaleB
Method
HOG DSIFT
NMI ARI F-score NMI ARI F-score
SAE10 93.43 82.57 83.07 87.54 75.82 76.50
DSC11 96.91 90.25 89.46 90.85 83.00 83.45
DDL 96.82 88.97 89.13 90.20 81.83 83.42
Joint DTL 98.93 93.43 92.06 93.26 85.62 85.86
10. S. Gao, Y. Zhang, K. Jia, J. Lu, and Y. Zhang, “Single sample face recognition via learning deep supervised autoencoders,”
IEEE Transactions on Information Forensics and Security, 2015.
11. X. Peng, J. Feng, S. Xiao, J. Lu, Z. Yi, and S. Yan, “Deep sparse subspace clustering,” arXiv preprint arXiv:1709.08374,
2017.
21. What is Inverse Problem?
Inverse problem is given by equation:
• Operator A defines the problem
• Denoising - identity
• Super-resolution - subsampling
• Deblurring - convolution
• Reconstruction - projection
!21
y = Ax +η
22. Sparsity based Solution
• Exploits the sparsity of the image in some domain.
• Assume that sparsifying basis is known (DCT, wavelet
etc.).
• where is sparse representation and is fixed basis.
• Are fixed basis the best possible option?
φ
!22
α
y = Ax +η = Aφα +η
23. Adaptive Learning based Solution
!23
Transform learning learns basis adaptively from the image
patches
Z =[z1
|z2
|...|zK
]
Data consistency Transform learning
Pi
x : ith patch of the image
min
x,T ,Z
|| y − Ax ||2
2
+λ( ||TPi
x − zi
||2
2
i
∑ + µ(||T ||F
2
−logdetT +γ || zi
||0
)
24. Proposed DTL Inversion
!24
DTL12 learns multiple levels of transforms. The problem is
formulated as
Data consistency Deep transform learning
Z =[z1
|z2
|...|zK
]Pi
x : ith patch of the image
min
x,T1,T2 ,T3,Z
|| y − Ax ||2
2
+λ( ||T3
i
∑ T2
T1
Pi
x − z ||F
2
+µ (||Ti
||F
2
−logdetTi
)
j=1
3
∑ +γ || zi
||1
)
12. J. Maggu and A. Majumdar ,”Transductive Inversion via Deep Transform Learning”, Signal Processing(submitted).
T1Pi x > 0
T2T1Pi x > 0
25. Deblurring Results
!25
Images Blurry RCSR13 GBD14 DeblurGAN15 Proposed
Baby 0.78 0.76 0.85 0.86 0.89
Bird 0.76 0.74 0.83 0.84 0.85
Butterfly 0.48 0.47 0.62 0.63 0.65
Head 0.66 0.65 0.72 0.84 0.84
Woman 0.73 0.71 0.80 0.80 0.82
Comparative Debluring Perfomance (SSIM)
13. M. Tofighi, Y. Li and V. Monga, "Blind Image Deblurring Using Row–Column Sparse Representations," IEEE Signal
Processing Letters, 2018.
14. Y. Bai, G. Cheung, X. Liu and W. Gao, "Graph-Based Blind Image Deblurring From a Single Photograph," IEEE
Transactions on Image Processing, 2019.
15. O. Kupyn, V. Budzan, M. Mykhailych, D. Mishkin and J. Matas, "DeblurGAN: Blind Motion Deblurring Using
Conditional Adversarial Networks," IEEE Conference on Computer Vision and Pattern Recognition, 2018.
26. Deblurring Results
!26
Man Left to Right: Original, Blurred image, RCSR, GBD, DeblurGAN and Proposed
Original Blurred RCSR GBD DeblurGAN Proposed
28. Supervised Deep Transform Learning
Label-Consistent DTL :
Multi-class classification
Multi-label classification16
!28
min
Ti
′s,Z ,M
||TN (φ...(T2 (φ(T1X))))− Z ||F
2
+λ (µ ||Ti ||F
2
−logdetTi )
i
∑ +η ||Q −φ(MZ)||F
2
supervision term
Deep TL
16. V. Singhal, J. Maggu and A. Majumdar, “Simultaneous Detection of Multiple Appliances from Smart-meter Measurements
via Multi-Label Consistent Deep Dictionary Learning and Deep Transform Learning,” IEEE Transactions on Smart Grid, 2019.
.
29. Multi-class Classification Results
!29
Technique YaleB AR
Stacked Denoising Autoencoder 42.81 37.60
Stacked Group Sparse Autoencoder 66.27 32.50
Stacked Label Consistent Autoencoder 86.22 85.21
Discriminative Deep Belief Network 60.34 38.20
LC-KSVD 90.80 87.67
DDL (unsupervised) 93.35 92.66
LC-DDL 94.57 96.50
DTL (unsupervised) 97.67 96.80
LCTL 1-layer 98.80 97.80
LCTL 2-layers 98.87 97.91
LCTL 3-layers 98.65 98.89
LCTL 4-layers 97.24 96.16
Classification Accuracy on AR and YaleB Face Recognition Datasets
30. NILM as Multi-label Classification
Problem
!30
Aggregated load
Supervised Deep
transform learning
A1
A2
An
Appliance states
31. Results on Energy Datasets
!31
Dataset
REDD dataset Pecan street dataset
Micro-F1 Macro-F1 Energy error Micro-F1 Macro-F1 Energy error
MLKNN 0.6034 0.5931 0.1067 0.6263 0.6227 0.0989
RAKEL 0.5749 0.5334 0.9948 0.6663 0.6620 0.9995
Proposed (1 layer) 0.5884 0.5838 0.0983 0.6079 0.6079 0.0236
Proposed (2 layers) 0.5905 0.5857 0.0892 0.6082 0.6089 0.0223
Proposed (3 layers) 0.6001 0.5981 0.0766 0.6104 0.6104 0.0115
Proposed (4 layers)
0.5914 0.5951 0.0827 0.6096 0.6087 0.0228
Performance on REDD and Pecan street Datasets
17. M.-L. Zhang and Z.-H. Zhou, “A k-nearest neighbor based algorithm for multi-label classification,” in Granular Computing, 2005.
18. G. Tsoumakas and I. Vlahavas, “Random k-labelsets: An ensemble method for multilabel classification,” Mach. Learn. ECML 2007.
17
18
33. Subspace Clustering
A special case of spectral clustering, where data samples
from same cluster are assumed to lie in same subspace.
• Each data point expressed as a linear combination of
others: ! with ! the i-th
sample, ! gathers all the other samples
column-wise, and ! states for the corresponding
linear weight vector.
• An affinity matrix ! is computed from the
! to quantify the similarity (inverse distance)
between the samples.
• The clusters are segmented by applying a cut technique
(eg, N-Cut).
(∀i ∈{1,...,n}) xi
= Xic ci xi
∈!m
Xic ∈!m×n−1
ci
∈!n−1
A ∈!n×n
(ci
)1≤i≤n
!33
34. Subspace Clustering
!34
Illustration of the subspace clustering16 framework based on sparse and
low-rank representation approaches for building the affinity matrix
19. A. Sobral, “Robust Low-rank and Sparse Decomposition for Moving Object Detection: From Matrices to Tensors,” 10.13140/RG.
2.2.33578.82884.
35. Transformed Subspace Clustering
!35
Illustration of the transformed subspace clustering framework based on
sparse and low-rank representation approaches for building the affinity matrix
On transformed
coefficient space
Joint solution
36. Deep Transformed Subspace Clustering
● Learn the linear weight vector on the transformed coefficient
space.
● Transformed locally linear manifold clustering22:
● Transformed sparse subspace clustering20,21:
● Transformed low rank subspace clustering20,21:
R(C) = 0
R(C) =|| C ||1
R(C) =|| C ||*
!36
min
T3,T2 ,T1,Z,C
||T3T2T1X − Z ||F
2
+λ (||Ti ||F
2
i=1
3
∑ − logdetTi )+γ || zi − Zic ci ||2
2
+R(C)
i
∑
20. J. Maggu, A. Majumdar and E. Chozenoux, “Transformed Subspace Clustering,” IEEE Transactions on Knowledge and Data Engineering (accepted).
21. J. Maggu, A. Majumdar, E. Chozenoux and G. Chierchia , “Deeply Transformed Subspace Clustering,” Signal Processing (major revision).
22. J. Maggu, A. Majumdar, and E. Chozenoux , “Transformed Locally Linear Manifold Clustering,” EUSIPCO 2018.
clustering termdeep transform
37. Experimental Results: EYALEB
!37
Method DSC23 DKM24 DMF25 DTLLMC DTSSC
Accuracy 88.00 91.00 89.00 93.13 99.26
NMI 0.90 0.92 0.90 0.92 0.95
ARI 0.83 0.90 0.83 0.91 0.97
Precision 0.79 0.91 0.80 0.94 0.99
F-Score 0.83 0.90 0.84 0.94 0.97
Comparison with benchmarks on EYALEB
23. X. Peng, S. Xiao, J. Feng, W. Y. Yau and Z. Yi, “Deep Sub-space Clustering with Sparsity Prior,” IJCAI, 2016.
24. B. Yang, X. Fu, N. D. Sidiropoulos and M. Hong, “Towards k-means-friendly spaces: Simultaneous deep learning and
clustering,” ICML, 2017.
25. G. Trigeorgis, K. Bousmalis, S. Zafeiriou and B. W. Schuller, "A Deep Matrix Factorization Method for Learning Attribute
Representations," IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017.
38. Experimental Results: EYALEB
!38
Method
DTSSC
1-layer
DTSSC
2-layers
DTSSC
3-layers
Accuracy 99.22 99.23 99.26
NMI 0.9448 0.9451 0.9476
ARI 0.9656 0.9663 0.9666
Precision 0.9887 0.9900 0.9912
F-Score 0.9567 0.9610 0.9667
Comparison with benchmarks on EYALEB
The joint formulation reduces the chances of masking relevant
clusters by noisy features and avoids the need for a preliminary
step of feature extraction.
40. Convolutional Transform Learning
● In standard TL, a dense basis is learnt.
● Proposal: Learn a set of independent filters that convolve on
images to learn representations.
● Motivation: The pivotal connection between CNNs and
CTL; but CTL is unexplored.
● Research gaps being addressed:
● Unlike CNNs, CTL works as unsupervised learning.
● Learnt filters are guaranteed to be mutually different.
● CNNs analysed via Convolutional sparse coding.
!40
41. Convolutional Transform Learning
● Input: Dataset: with M entries in
● Proposed model:
T convolutive transform, which gathers a set of K
kernels, i.e.
Toeplitz matrix such that
A matrix of coefficients associated to each
entry of the dataset.
● Goal: Estimate dense filters and sparse coefficients
from the
︎
m ∈ 1,…,M{ }
{tm}1≤m≤M
{Zm}1≤m≤M {x(m)
}1≤m≤M
!41
!"
T = [t1 |...| tK ]∈!K×K
{x(m)
}1≤m≤M
(∀m ∈{1,...,M}),
χ(M )
T ≈ Zm
(χ(m)
)1≤m≤K ∈!N×K
χ(m)
T = [t1 ∗ x(m)
|...| tK ∗ x(m)
]
Zm = [z1
(m)
|...| ZK
(m)
]
42. Convolutional Transform Learning
Learns convolved features in an unsupervised way
!
with
min
T ,Z
1
2
||T ∗ X(m)
− Zm ||F
2
m=1
M
∑ + µ ||T ||F
2
−λ log |T | +β || Z ||1 +ι[0,+∞[NM ×K (Z)
!42
26. J. Maggu, A. Majumdar and E. Chozenoux, “Convolutional Transform Learning,” IEEE ICONIP 2018.
Z = [Z1
⊤
|…| ZM
⊤
]⊤
∈!NM ×K
.
min
T ,Z
1
2
|| χ(m)
T − Zm ||F
2
m=1
M
∑ + µ ||T ||F
2
−λ log |T | +β || Z ||1 +ι[0,+∞[NM ×K (Z)
46. Semi-coupled transform learning
!46
Coefficients Z1
Coefficients Z2
Transform T2
Transform T1
Common feature space
Synthesis
LR image
HR image
Photo
Sketch
Source view action
Target view action
Resolution
Data X1
Data X2
Projection
47. Semi-coupled transform learning
• Comparison of heterogeneous
samples
• Data can be from different
sources
• Eg. face sketch and photo for
matching
!47
T1 T2
Z2Z1
X1 X2
M
53. Image super-resolution Results
!53
Image name Lena Barbara Pepper Cameraman
Color CDL 30.79 28.21 29.76 27.86
Proposed 33.03 30.28 31.81 30.14
Gray scale CDL 31.27 28.98 30.46 28.70
Proposed 34.55 31.17 32.68 30.85
PSNR for super-resolution
54. Cross lingual document retrieval results
!54
Algorithm
Europarl Wikipedia
Accuracy MRR Accuracy MRR
OPCA28 97.42 0.9846 72.55 0.7734
CPLSA28 97.16 0.9782 45.79 0.5130
CDL29 98.12 0.9839 72.79 28.70
Proposed 99.54 0.9896 78.68 0.8002
Comparable document retrieval
28. Platt J.C., Toutanova, “K.:Association for Computational Linguistics", Conference on Empirical Methods in Natural
Language Processing, 2011.
29. Mehrotra R., Chu D., Haider S.A., Kakadiaris I.A, “Towards Learning Coupled Representations for Cross-Lingual
Information Retrieval”.
OPCA: Oriented Principal Component Analysis
CPLSA: Coupled Probabilistic Latent Semantic Analysis
CDL: Coupled dictionary learning
MRR: Mean Reciprocal Rank
55. Outline
●Supervised TL
●Unsupervised DTL
●Supervised DTL
●Deep Transformed Subspace Clustering
●Convolutional TL
●Semi-coupled TL
●Future Work
● Deeply Coupled TL
● Deep Transform Information Fusion Network
!55
56. Deeply-coupled transform learning
!56
Coefficients Z1
Coefficients Z2
Deep Transform T2
Deep Transform T1
Common feature space
Synthesis
LR image
HR image
Photo
Sketch
Source view action
Target view action
Resolution
Data X2
Projection
Data X1
57. Deeply Coupled transform learning
• Comparison of heterogeneous
samples
• Data can be from different
sources
• Eg. face sketch and photo for
matching
!57
X1 X2
Z2Z1
M
T11
T12 T22
T21
59. Deep Transform Information Fusion Network
● The network learns if the two
inputs (images) presented are
related or not.
● Eg. Verification Task
!59
Architecture 1
60. Deep Transform Information Fusion Network
● The network learns if the two
inputs (images) presented are
related or not.
● Eg. Verification Task
!60
Architecture 2
61. Deep Transform Information Fusion Network
● The network learns if the two
inputs (images) presented are
related or not.
● Eg. Verification Task
!61
Architecture 3
62. Publications(Journals)
!62
1. J. Maggu, A. Majumdar and E. Chouzenoux, “Transformed Subspace Clustering”, IEEE Transactions on Knowledge
and Data Engineering (accepted).
2. J. Maggu, H. Agarwal and A. Majumdar, “Label Consistent Transform Learning for Hyperspectral Image
Classification”, IEEE Geosciences and Remote Sensing Letters, Vol. 16 (9), pp. 1502-1506, 2019
3. V. Singhal, J. Maggu and A. Majumdar, “Simultaneous Detection of Multiple Appliances from Smart-meter
Measurements via Multi-Label Consistent Deep Dictionary Learning and Deep Transform Learning” IEEE
Transactions on Smart Grid, Vol. 10 (3), pp. 2969-2978, 2019.
4. J. Maggu, P. Singh and A. Majumdar, “Multi-echo Reconstruction from Partial K-space Scans via Adaptively Learnt
Basis”, Magnetic Resonance Imaging, Vol. 45, pp. 105-112, 2018.
5. J. Maggu and A. Majumdar, “Kernel Transform Learning”, Pattern Recognition Letters, Vol. 117, pp. 117-122, 2017.
6. J. Maggu, A. Majumdar, E. Chouzenoux and G. Chierchia, “Deeply Transformed Subspace Clustering”, Signal
Processing (major revision).
7. J. Maggu and A. Majumdar, “Dynamic MRI Reconstruction with Deep Transform Learning Prior”, Magnetic
Resonance Imaging, (major revision)
8. J. Maggu and A. Majumdar, “Transductive Inversion via Deep Transform Learning”, Signal Processing (submitted).
63. Publications(Conferences)
!63
1. J. Maggu and A. Majumdar, “Supervised Kernel Transform Learning”, IEEE IJCNN 2019.
2. J. Maggu, E. Chouzenoux, G. Chierchia and A. Majumdar, “Convolutional Transform Learning”, ICONIP,
pp. 162-174, 2018.
3. J. Maggu and A. Majumdar, "Semi-Coupled Transform Learning", ICONIP, pp. 141-150, 2018.
4. J. Maggu, A. Majumdar and E. Chouzenoux, “Transformed Locally Linear Manifold Clustering”,
EUSIPCO, pp. 1057-1061, 2018.
5. J. Maggu and A. Majumdar, "Unsupervised Deep Transform Learning", IEEE ICASSP, pp. 6782-6786,
2018.
6. J. Maggu, R. Hussein, A. Majumdar and R. Ward, "Impulse Denoising via Transform Learning", IEEE
GlobalSIP, pp. 1250-1254, 2017.
7. J. Maggu and A. Majumdar, “Greedy Deep Transform Learning”, IEEE ICIP, pp. 1822-1826, 2017.
8. J. Maggu and A. Majumdar, “Robust Transform Learning”, IEEE ICASSP, pp. 1467-1471, 2017.
9. J. Maggu and A. Majumdar, "Alternate Formulation for Transform Learning", ICVGIP, pp. 501-508, 2016.