DOCSLIB.ORG

Dataless Model Selection with the Deep Frame Potential

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Dataless Model Selection with the Deep Frame Potential Dataless Model Selection with the Deep Frame Potential Calvin Murdock1 Simon Lucey1,2 1Carnegie Mellon University 2Argo AI fcmurdock,[email protected] Abstract Choosing a deep neural network architecture is a funda- mental problem in applications that require balancing per- formance and parameter efficiency. Standard approaches (a) Chain Network (b) Residual Network (c) Densely Connected rely on ad-hoc engineering or computationally expensive Convolutional Network validation on a specific dataset. We instead attempt to quantify networks by their intrinsic capacity for unique and robust representations, enabling efficient architecture com- ≈ ≈ ≈ parisons without requiring any data. Building upon theoret- ical connections between deep learning and sparse approx- imation, we propose the deep frame potential: a measure (d) Induced Dictionary Structures for Sparse Approximation of coherence that is approximately related to representation stability but has minimizers that depend only on network Figure 1: Why are some deep neural network architectures better structure. This provides a framework for jointly quantifying than others? In comparison to (a) standard chain connections, skip the contributions of architectural hyper-parameters such as connections like those in (b) ResNets [13] and (c) DenseNets [15] have demonstrated significant improvements in training effective- depth, width, and skip connections. We validate its use as ness, parameter efficiency, and generalization performance. We a criterion for model selection and demonstrate correlation provide one possible explanation for this phenomenon by approx- with generalization error on a variety of common residual imating network activations as (d) solutions to sparse approxima- and densely connected network architectures. tion problems with different induced dictionary structures. 1. Introduction low machine learning techniques like support vector ma- Deep neural networks have dominated nearly every chines [6] were aided by theoretical tools like the VC- benchmark within the field of computer vision. While this dimension [31] for determining when their predictions modern influx of deep learning originally began with the could be trusted to avoid overfitting. Deep neural networks, task of large-scale image recognition [18], new datasets, on the other hand, have eschewed similar analyses due to loss functions, and network configurations have quickly ex- their complexity. Theoretical explorations of deep network arXiv:2003.13866v1 [cs.LG] 30 Mar 2020 panded its scope to include a much wider range of appli- generalization [24] are often disconnected from practical cations. Despite this, the underlying architectures used to applications and rarely provide actionable insight into how learn effective image representations are generally consis- architectural hyper-parameters contribute to performance. tent across all of them. This can be seen through the com- Building upon recent connections between deep learn- munity’s quick adoption of the newest state-of-the-art deep ing and sparse approximation [26, 23], we instead interpret networks from AlexNet [18] to VGGNet [28], ResNets [13], feed-forward deep networks as algorithms for approximate DenseNets [15], and so on. But this begs the question: why inference in related sparse coding problems. These prob- do some deep network architectures work better than oth- lems aim to optimally reconstruct zero-padded input images ers? Despite years of groundbreaking empirical results, an as sparse, nonnegative linear combinations of atoms from answer to this question still remains elusive. architecture-dependent dictionaries, as shown in Fig.1. We Fundamentally, the difficulty in comparing network ar- propose to indirectly analyze practical deep network archi- chitectures arises from the lack of a theoretical founda- tectures with complicated skip connections, like residual tion for characterizing their generalization capacities. Shal- networks (ResNets) [13] and densely connected convolu- tional networks (DenseNets) [15], simply through the dic- tionary structures that they induce. To accomplish this, we introduce the deep frame poten- = tial for summarizing the interactions between parameters in feed-forward deep networks. As a lower bound on mutual (a) Chain Network Gram Matrix (b) ResNet (c) DenseNet coherence–the maximum magnitude of the normalized in- 0.035 0.4 ner products between all pairs of dictionary atoms [9]–it Dense Network Dense Network 0.03 Residual Network Residual Network Chain Network Chain Network 0.3 is theoretically tied to generalization properties of the re- 0.025 lated sparse coding problems. However, its minimizers de- 0.02 0.2 pend only on the dictionary structures induced by the cor- 0.015 0.01 responding network architectures. This enables dataless Validation Error 0.1 0.005 model comparison by jointly quantifying contributions of Minimum Frame Potential 0 0 depth, width, and connectivity. 103 104 105 106 107 103 104 105 106 107 Number of Parameters (Log Scale) Number of Parameters (Log Scale) Our approach is motivated by sparse approximation the- (d) Minimum Deep Frame Potential (e) Validation Error ory [11], a field that encompasses properties like uniqueness and robustness of shallow, overcomplete representations. In Figure 2: Parameter count is not a good indicator of generaliza- sparse coding, capacity is controlled by the number of dic- tion performance for deep networks. Instead, we compare differ- tionary atoms used in sparse data reconstructions. While ent network architectures via the minimum deep frame potential, more parameters allow for more accurate representations, the average nonzero magnitude of inner products between atoms of they may also increase input sensitivity for worse general- architecture-induced dictionaries. In comparison to (a) chain net- ization performance. Conceptually, this is comparable to works, skip connections in (b) residual networks and (c) densely overfitting in nearest-neighbor classification, where repre- connected networks produce Gram matrix structures with more nonzero elements allowing for (d) lower deep frame potentials sentations are sparse, one-hot indicator vectors correspond- across network sizes. This correlates with improved parameter ing to nearest training examples. As the number of train- efficiency giving (e) lower validation error with fewer parameters. ing data increases, the distance between them decreases, so they are more likely to be confused with one another. Simi- larly, nearby dictionary atoms may introduce instability that tional networks with skip connections, of which ResNets causes representations of similar data points to become very and DenseNets are shown to be special cases. Further- far apart leading to poor generalization performance. Thus, more, we derive an analytic expression for the minimum there is a fundamental tradeoff between the capacity and ro- value in the case of fully-connected chain networks. Exper- bustness of shallow representations due to the proximity of imentally, we demonstrate correlation with validation error dictionary atoms as measured by mutual coherence. across a variety of network architectures. However, deep representations have not shown the same correlation between model size and sensitivity [34]. While 2. Background and Related Work adding more layers to a deep neural network increases its capacity, it also simultaneously introduces implicit regular- Due to the vast space of possible deep network archi- ization to reduce overfitting. This can be explained through tectures and the computational difficulty in training them, the proposed connection to sparse coding, where additional deep model selection has largely been guided by ad-hoc layers increase both capacity and effective input dimension- engineering and human ingenuity. While progress slowed ality. In a higher-dimensional space, dictionary atoms can in the years following early breakthroughs [20], recent in- be spaced further apart for more robust representations. Fur- terest in deep learning architectures began anew due to thermore, architectures with denser skip connections induce empirical successes largely attributed to computational ad- dictionary structures with more nonzero elements, which vances like efficient training using GPUs and rectified lin- provides additional freedom to further reduce mutual co- ear unit (ReLU) activation functions [18]. Since then, nu- herence with fewer parameters as shown in Fig.2. merous architectural changes have been proposed. For ex- We propose to use the minimum deep frame potential as ample, much deeper networks with residual connections a cue for model selection. Instead of requiring expensive were shown to achieve consistently better performance with validation on a specific dataset to approximate generaliza- fewer parameters [13]. Building upon this, densely con- tion performance, architectures are chosen based on how nected convolutional networks with skip connections be- efficiently they can reduce the minimum achievable mutual tween more layers yielded even better performance [15]. coherence with respect to the number of model parame- While theoretical explanations for these improvements were ters. In this paper, we provide an efficient frame poten- lacking, consistent experimentation on standardized bench- tial minimization method for a general class of convolu- mark datasets continued to drive empirical success. T T However, due to slowing progress and the need for in- network f(x) = φl(Bl ··· φ1(B1 (x))) constructed
Load more

Recommended publications
  • Accelerating Matching Pursuit for Multiple Time-Frequency Dictionaries
    Proceedings of the 23rd International Conference on Digital Audio Effects (DAFx-20), Vienna, Austria, September 8–12, 2020 ACCELERATING MATCHING PURSUIT FOR MULTIPLE TIME-FREQUENCY DICTIONARIES ZdenˇekPr˚uša,Nicki Holighaus and Peter Balazs ∗ Acoustics Research Institute Austrian Academy of Sciences Vienna, Austria [email protected],[email protected],[email protected] ABSTRACT An overview of greedy algorithms, a class of algorithms MP Matching pursuit (MP) algorithms are widely used greedy meth- falls under, can be found in [10, 11] and in the context of audio ods to find K-sparse signal approximations in redundant dictionar- and music processing in [12, 13, 14]. Notable applications of MP ies. We present an acceleration technique and an implementation algorithms in the audio domain include analysis [15], [16], coding of the matching pursuit algorithm acting on a multi-Gabor dictio- [17, 18, 19], time scaling/pitch shifting [20] [21], source separation nary, i.e., a concatenation of several Gabor-type time-frequency [22], denoising [23], partial and harmonic detection and tracking dictionaries, consisting of translations and modulations of possi- [24]. bly different windows, time- and frequency-shift parameters. The We present a method for accelerating MP-based algorithms proposed acceleration is based on pre-computing and thresholding acting on a single Gabor-type time-frequency dictionary or on a inner products between atoms and on updating the residual directly concatenation of several Gabor dictionaries with possibly different in the coefficient domain, i.e., without the round-trip to thesig- windows and parameters. The main idea of the present accelera- nal domain.
    [Show full text]
  • Paper, We Present Data Fusion Across Multiple Signal Sources
    68 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 51, NO. 1, JANUARY 2016 A Configurable 12–237 kS/s 12.8 mW Sparse-Approximation Engine for Mobile Data Aggregation of Compressively Sampled Physiological Signals Fengbo Ren, Member, IEEE, and Dejan Markovic,´ Member, IEEE Abstract—Compressive sensing (CS) is a promising technology framework, the CS framework has several intrinsic advantages. for realizing low-power and cost-effective wireless sensor nodes First, random encoding is a universal compression method that (WSNs) in pervasive health systems for 24/7 health monitoring. can effectively apply to all compressible signals regardless of Due to the high computational complexity (CC) of the recon- struction algorithms, software solutions cannot fulfill the energy what their sparse domain is. This is a desirable merit for the efficiency needs for real-time processing. In this paper, we present data fusion across multiple signal sources. Second, sampling a 12—237 kS/s 12.8 mW sparse-approximation (SA) engine chip and compression can be performed at the same stage in CS, that enables the energy-efficient data aggregation of compressively allowing for a sampling rate that is significantly lower than the sampled physiological signals on mobile platforms. The SA engine Nyquist rate. Therefore, CS has a potential to greatly impact chip integrated in 40 nm CMOS can support the simultaneous reconstruction of over 200 channels of physiological signals while the data acquisition devices that are sensitive to cost, energy consuming <1% of a smartphone’s power budget. Such energy- consumption, and portability, such as wireless sensor nodes efficient reconstruction enables two-to-three times energy saving (WSNs) in mobile and wearable applications [5].
    [Show full text]
  • Improved Greedy Algorithms for Sparse Approximation of a Matrix in Terms of Another Matrix
    Improved Greedy Algorithms for Sparse Approximation of a Matrix in terms of Another Matrix Crystal Maung Haim Schweitzer Department of Computer Science Department of Computer Science The University of Texas at Dallas The University of Texas at Dallas Abstract We consider simultaneously approximating all the columns of a data matrix in terms of few selected columns of another matrix that is sometimes called “the dic- tionary”. The challenge is to determine a small subset of the dictionary columns that can be used to obtain an accurate prediction of the entire data matrix. Previ- ously proposed greedy algorithms for this task compare each data column with all dictionary columns, resulting in algorithms that may be too slow when both the data matrix and the dictionary matrix are large. A previously proposed approach for accelerating the run time requires large amounts of memory to keep temporary values during the run of the algorithm. We propose two new algorithms that can be used even when both the data matrix and the dictionary matrix are large. The first algorithm is exact, with output identical to some previously proposed greedy algorithms. It takes significantly less memory when compared to the current state- of-the-art, and runs much faster when the dictionary matrix is sparse. The second algorithm uses a low rank approximation to the data matrix to further improve the run time. The algorithms are based on new recursive formulas for computing the greedy selection criterion. The formulas enable decoupling most of the compu- tations related to the data matrix from the computations related to the dictionary matrix.
    [Show full text]
  • Privacy Preserving Identification Using Sparse Approximation With
    Privacy Preserving Identification Using Sparse Approximation with Ambiguization Behrooz Razeghi, Slava Voloshynovskiy, Dimche Kostadinov and Olga Taran Stochastic Information Processing Group, Department of Computer Science, University of Geneva, Switzerland behrooz.razeghi, svolos, dimche.kostadinov, olga.taran @unige.ch f g Abstract—In this paper, we consider a privacy preserving en- Owner Encoder Public Storage coding framework for identification applications covering biomet- +1 N M λx λx L M rics, physical object security and the Internet of Things (IoT). The X × − A × ∈ X 1 ∈ A proposed framework is based on a sparsifying transform, which − X = x (1) , ..., x (m) , ..., x (M) a (m) = T (Wx (m)) n A = a (1) , ..., a (m) , ..., a (M) consists of a trained linear map, an element-wise nonlinearity, { } λx { } and privacy amplification. The sparsifying transform and privacy L p(y (m) x (m)) Encoder | amplification are not symmetric for the data owner and data user. b Schematic Decoding List Probe +1 We demonstrate that the proposed approach is closely related (Private Decoding) y = x (m) + z d (a (m) , b) γL λy λy ≤ − to sparse ternary codes (STC), a recent information-theoretic 1 p (positions) − ´x y = ´x (Pubic Decoding) 1 m M (y) concept proposed for fast approximate nearest neighbor (ANN) ≤ ≤ L Data User b = Tλy (Wy) search in high dimensional feature spaces that being machine learning in nature also offers significant benefits in comparison Fig. 1: Block diagram of the proposed model. to sparse approximation and binary embedding approaches. We demonstrate that the privacy of the database outsourced to a for example biometrics, which being disclosed once, do not server as well as the privacy of the data user are preserved at a represent any more a value for the related security applications.
    [Show full text]
  • Column Subset Selection Via Sparse Approximation of SVD
    Column Subset Selection via Sparse Approximation of SVD A.C¸ivrila,∗, M.Magdon-Ismailb aMeliksah University, Computer Engineering Department, Talas, Kayseri 38280 Turkey bRensselaer Polytechnic Institute, Computer Science Department, 110 8th Street Troy, NY 12180-3590 USA Abstract Given a real matrix A 2 Rm×n of rank r, and an integer k < r, the sum of the outer products of top k singular vectors scaled by the corresponding singular values provide the best rank-k approximation Ak to A. When the columns of A have specific meaning, it might be desirable to find good approximations to Ak which use a small number of columns of A. This paper provides a simple greedy algorithm for this problem in Frobenius norm, with guarantees on( the performance) and the number of columns chosen. The algorithm ~ k log k 2 selects c columns from A with c = O ϵ2 η (A) such that k − k ≤ k − k A ΠC A F (1 + ϵ) A Ak F ; where C is the matrix composed of the c columns, ΠC is the matrix projecting the columns of A onto the space spanned by C and η(A) is a measure related to the coherence in the normalized columns of A. The algorithm is quite intuitive and is obtained by combining a greedy solution to the generalization of the well known sparse approximation problem and an existence result on the possibility of sparse approximation. We provide empirical results on various specially constructed matrices comparing our algorithm with the previous deterministic approaches based on QR factorizations and a recently proposed randomized algorithm.
    [Show full text]
  • Modified Sparse Approximate Inverses (MSPAI) for Parallel
    Technische Universit¨atM¨unchen Zentrum Mathematik Modified Sparse Approximate Inverses (MSPAI) for Parallel Preconditioning Alexander Kallischko Vollst¨andiger Abdruck der von der Fakult¨atf¨ur Mathematik der Technischen Universit¨at M¨unchen zur Erlangung des akademischen Grades eines Doktors der Naturwissenschaften (Dr. rer. nat.) genehmigten Dissertation. Vorsitzender: Univ.-Prof. Dr. Peter Rentrop Pr¨ufer der Dissertation: 1. Univ.-Prof. Dr. Thomas Huckle 2. Univ.-Prof. Dr. Bernd Simeon 3. Prof. Dr. Matthias Bollh¨ofer, Technische Universit¨atCarolo-Wilhelmina zu Braunschweig (schriftliche Beurteilung) Die Dissertation wurde am 15.11.2007 bei der Technischen Universit¨ateingereicht und durch die Fakult¨atf¨urMathematik am 18.2.2008 angenommen. ii iii Abstract The solution of large sparse and ill-conditioned systems of linear equations is a central task in numerical linear algebra. Such systems arise from many applications like the discretiza- tion of partial differential equations or image restoration. Herefore, Gaussian elimination or other classical direct solvers can not be used since the dimension of the underlying co- 3 efficient matrices is too large and Gaussian elimination is an O n algorithm. Iterative solvers techniques are an effective remedy for this problem. They allow to exploit sparsity, bandedness, or block structures, and they can be parallelized much easier. However, due to the matrix being ill-conditioned, convergence becomes very slow or even not be guaranteed at all. Therefore, we have to employ a preconditioner. The sparse approximate inverse (SPAI) preconditioner is based on Frobenius norm mini- mization. It is a well-established preconditioner, since it is robust, flexible, and inherently parallel. Moreover, SPAI captures meaningful sparsity patterns automatically.
    [Show full text]
  • A New Algorithm for Non-Negative Sparse Approximation Nicholas Schachter
    A New Algorithm for Non-Negative Sparse Approximation Nicholas Schachter To cite this version: Nicholas Schachter. A New Algorithm for Non-Negative Sparse Approximation. 2020. hal- 02888300v1 HAL Id: hal-02888300 https://hal.archives-ouvertes.fr/hal-02888300v1 Preprint submitted on 2 Jul 2020 (v1), last revised 9 Jun 2021 (v5) HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. A New Algorithm for Non-Negative Sparse Approximation Nicholas Schachter July 2, 2020 Abstract In this article we introduce a new algorithm for non-negative sparse approximation problems based on a combination of the approaches used in orthogonal matching pursuit and basis de-noising pursuit towards solving sparse approximation problems. By taking advantage of structural properties inherent to non-negative sparse approximation problems, a branch and bound (BnB) scheme is developed that enables fast and accurate recovery of underlying dictionary atoms, even in the presence of noise. Detailed analysis of the performance of the algorithm is discussed, with attention specically paid to situations in which the algorithm will perform better or worse based on the properties of the dictionary and the required sparsity of the solution.
    [Show full text]
  • Efficient Implementation of the K-SVD Algorithm Using
    E±cient Implementation of the K-SVD Algorithm using Batch Orthogonal Matching Pursuit Ron Rubinstein¤, Michael Zibulevsky¤ and Michael Elad¤ Abstract The K-SVD algorithm is a highly e®ective method of training overcomplete dic- tionaries for sparse signal representation. In this report we discuss an e±cient im- plementation of this algorithm, which both accelerates it and reduces its memory consumption. The two basic components of our implementation are the replacement of the exact SVD computation with a much quicker approximation, and the use of the Batch-OMP method for performing the sparse-coding operations. - Technical Report CS-2008-08.revised 2008 Batch-OMP, which we also present in this report, is an implementation of the Orthogonal Matching Pursuit (OMP) algorithm which is speci¯cally optimized for sparse-coding large sets of signals over the same dictionary. The Batch-OMP imple- mentation is useful for a variety of sparsity-based techniques which involve coding large numbers of signals. In the report, we discuss the Batch-OMP and K-SVD implementations and analyze their complexities. The report is accompanied by Matlabr toolboxes which implement these techniques, and can be downloaded at http://www.cs.technion.ac.il/~ronrubin/software.html. 1 Introduction Sparsity in overcomplete dictionaries is the basis for a wide variety of highly e®ective signal and image processing techniques. The basic model suggests that natural signals can be e±ciently explained as linear combinations of prespeci¯ed atom signals, where the linear coe±cients are sparse (most of them zero). Formally, if x is a column signal and D is the dictionary (whose columns are the atom signals), the sparsity assumption can be described by the following sparse approximation problem, 2 γ^ = Argmin γ 0 Subject To x Dγ ² : (1.1) γ k k k ¡ k2 · Technion - Computer Science Department In this formulation, γ is the sparse representation of x, ² the error tolerance, and k ¢ k0 is the `0 pseudo-norm which counts the non-zero entries.
    [Show full text]
  • Regularized Dictionary Learning for Sparse Approximation
    16th European Signal Processing Conference (EUSIPCO 2008), Lausanne, Switzerland, August 25-29, 2008, copyright by EURASIP REGULARIZED DICTIONARY LEARNING FOR SPARSE APPROXIMATION M. Yaghoobi, T. Blumensath, M. Davies Institute for Digital Communications, Joint Research Institute for Signal and Image Processing, University of Edinburgh, UK ABSTRACT keeping the dictionary fixed. This is followed by a second step in Sparse signal models approximate signals using a small number of which the sparse coefficients are kept fixed and the dictionary is elements from a large set of vectors, called a dictionary. The suc- optimized. This algorithm runs for a specific number of alternating cess of such methods relies on the dictionary fitting the signal struc- optimizations or until a specific approximation error is reached. The ture. Therefore, the dictionary has to be designed to fit the signal proposed method is based on such an alternating optimization (or class of interest. This paper uses a general formulation that allows block-relaxed optimization) method with some advantages over the the dictionary to be learned form the data with some a priori in- current methods in the condition and speed of convergence. formation about the dictionary. In this formulation a universal cost If the set of training samples is {y(i) : 1 ≤ i ≤ L}, where L function is proposed and practical algorithms are presented to min- is the number of training vectors, then sparse approximations are imize this cost under different constraints on the dictionary. The often found (for all i : 1 ≤ i ≤ L ) by, proposed methods are compared with previous approaches using (i) (i) 2 p synthetic and real data.
    [Show full text]
  • Structured Compressed Sensing - Using Patterns in Sparsity
    Structured Compressed Sensing - Using Patterns in Sparsity Johannes Maly Technische Universit¨atM¨unchen, Department of Mathematics, Chair of Applied Numerical Analysis [email protected] CoSIP Workshop, Berlin, Dezember 9, 2016 Overview Classical Compressed Sensing Structures in Sparsity I - Joint Sparsity Structures in Sparsity II - Union of Subspaces Conclusion Johannes Maly Structured Compressed Sensing - Using Patterns in Sparsity 2 of 44 Classical Compressed Sensing Overview Classical Compressed Sensing Structures in Sparsity I - Joint Sparsity Structures in Sparsity II - Union of Subspaces Conclusion Johannes Maly Structured Compressed Sensing - Using Patterns in Sparsity 3 of 44 Classical Compressed Sensing Compressed Sensing N Let x 2 R be some unknown k-sparse signal. Then, x can be recovered from few linear measurements y = A · x m×N m where A 2 R is a (random) matrix, y 2 R is the vector of measurements and m N. Johannes Maly Structured Compressed Sensing - Using Patterns in Sparsity 4 of 44 Classical Compressed Sensing Compressed Sensing N Let x 2 R be some unknown k-sparse signal. Then, x can be recovered from few linear measurements y = A · x m×N m where A 2 R is a (random) matrix, y 2 R is the vector of measurements and m N. It is sufficient to have N m Ck log & k measurements to recover x (with high probability) by greedy strategies, e.g. Orthogonal Matching Pursuit, or convex optimization, e.g. `1-minimization. Johannes Maly Structured Compressed Sensing - Using Patterns in Sparsity 5 of 44 OMP INPUT: matrix A; measurement vector y: INIT: T0 = ;; x0 = 0: ITERATION: until stopping criterion is met T jn+1 arg maxj2[N] (A (y − Axn))j ; Tn+1 Tn [ fjn+1g; xn+1 arg minz2RN fky − Azk2; supp(z) ⊂ Tn+1g : OUTPUT: then ~-sparse approximationx ^ := xn~ Classical Compressed Sensing Orthogonal Matching Pursuit OMP is a simple algorithm that tries to find the true support of x by k greedy steps.
    [Show full text]
  • Sparse Estimation for Image and Vision Processing
    Sparse Estimation for Image and Vision Processing Julien Mairal Inria, Grenoble SPARS summer school, Lisbon, December 2017 Julien Mairal Sparse Estimation for Image and Vision Processing 1/187 Course material (freely available on arXiv) J. Mairal, F. Bach and J. Ponce. Sparse Modeling for Image and Vision Processing. Foundations and Trends in Computer Graphics and Vision. 2014. F. Bach, R. Jenatton, J. Mairal, and G. Obozinski. Optimization with sparsity-inducing penalties. Foundations and Trends in Machine Learning, 4(1). 2012. Julien Mairal Sparse Estimation for Image and Vision Processing 2/187 Outline 1 A short introduction to parsimony 2 Discovering the structure of natural images 3 Sparse models for image processing 4 Optimization for sparse estimation 5 Application cases Julien Mairal Sparse Estimation for Image and Vision Processing 3/187 Part I: A Short Introduction to Parcimony Julien Mairal Sparse Estimation for Image and Vision Processing 4/187 1 A short introduction to parsimony Early thoughts Sparsity in the statistics literature from the 60’s and 70’s Wavelet thresholding in signal processing from 90’s The modern parsimony and the ℓ1-norm Structured sparsity Compressed sensing and sparse recovery 2 Discovering the structure of natural images 3 Sparse models for image processing 4 Optimization for sparse estimation 5 Application cases Julien Mairal Sparse Estimation for Image and Vision Processing 5/187 Early thoughts (a) Dorothy Wrinch (b) Harold Jeffreys 1894–1980 1891–1989 The existence of simple laws is, then, apparently, to be regarded as a quality of nature; and accordingly we may infer that it is justifiable to prefer a simple law to a more complex one that fits our observations slightly better.
    [Show full text]
  • Exact Sparse Approximation Problems Via Mixed-Integer Programming: Formulations and Computational Performance
    Exact Sparse Approximation Problems via Mixed-Integer Programming: Formulations and Computational Performance Sébastien Bourguignon, Jordan Ninin, Hervé Carfantan, Marcel Mongeau To cite this version: Sébastien Bourguignon, Jordan Ninin, Hervé Carfantan, Marcel Mongeau. Exact Sparse Approxi- mation Problems via Mixed-Integer Programming: Formulations and Computational Performance. IEEE Transactions on Signal Processing, Institute of Electrical and Electronics Engineers, 2016, 64 (6), pp.1405-1419. 10.1109/TSP.2015.2496367. hal-01254856 HAL Id: hal-01254856 https://hal.archives-ouvertes.fr/hal-01254856 Submitted on 12 Jan 2016 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. 1 Exact Sparse Approximation Problems via Mixed-Integer Programming: Formulations and Computational Performance Sebastien´ Bourguignon, Jordan Ninin, Herve´ Carfantan, and Marcel Mongeau, Member, IEEE Abstract—Sparse approximation addresses the problem of too difficult in practical large-scale instances. Indeed, the approximately fitting a linear model with a solution having as few Q brute-force approach that amounts to exploring all the K non-zero components as possible. While most sparse estimation possible combinations, is computationally prohibitive. In the algorithms rely on suboptimal formulations, this work studies the abundant literature on sparse approximation, much work has performance of exact optimization of `0-norm-based problems through Mixed-Integer Programs (MIPs).
    [Show full text]