Hierarchical approximations for data reduction and learning at multiple scales

Prashant Shekhar; Abani Patra

doi:10.3934/fods.2020008

Article Contents

2020, Volume 2, Issue 2: 123-154. Doi: 10.3934/fods.2020008

This issue Previous Article Topological reconstruction of sub-cellular motion with Ensemble Kalman velocimetry Next Article A Bayesian nonparametric test for conditional independence

Hierarchical approximations for data reduction and learning at multiple scales

Prashant Shekhar^, and
Abani Patra

Data Intensive Studies Center, Tufts University, Medford, MA, 02155, USA

^* Corresponding author: Prashant Shekhar
^* Corresponding author: Prashant Shekhar

Published: July 2020

This work was funded by the grants NSF1821311, NSF1645053, NSF1621853

Abstract / Introduction Full Text(HTML) Figure(14) / Table(2) Related Papers Cited by

Abstract

This paper describes a hierarchical learning strategy for generating sparse representations of multivariate datasets. The hierarchy arises from approximation spaces considered at successively finer scales. A detailed analysis of stability, convergence and behavior of error functionals associated with the approximations are presented, along with a well chosen set of applications. Results show the performance of the approach as a data reduction mechanism for both synthetic (univariate and multivariate) and a real dataset (geo-spatial). The sparse representation generated is shown to efficiently reconstruct data and minimize error in prediction. The approach is also shown to generalize well to unseen samples, extending its prospective application to statistical learning problems.

Keywords:

Mathematics Subject Classification: Primary:68W25, 65D15;Secondary:65D12.

Citation:

Full Text(HTML)

Figure 4. Scalewise performance of Algorithm 1 on TF2. The plots in the left column((a), (c), (e) and (g)) show the density of basis functions at each scale, selected while identifying the sparse representation. Rank here refers to the cardinality of basis function set $ B^s $. The right column ((b), (d), (f) and (h)) shows the corresponding scalewise reconstruction of the underlying function. The small dots show the samples from true function while star markers represent the sparse representation $ D^s_{sparse} $. The solid curve is the reconstruction from the sparse representation. The fraction in bracket (for example (11/200) for (b)) shows the number of data points chosen out of 200 as $ D^s_{sparse} $

Download: Full-size image PowerPoint slide

Figure 1. Univariate {(a): Gramancy and Lee function - Test function 1 (TF1); (b): 1-D Schwefel function - TF2} and multivariate {(c): Dropwave function - TF3; (d): 2-D Schwefel function - TF4} test functions considered for studying the performance of Algorithm 1

Download: Full-size image PowerPoint slide

Figure 2. Convergence behavior on the test functions measured in 2-norm prediction error on the observed data. Top row ((a) and (b)) shows the performance on univariate functions with bottom row for multivariate functions ((c) and (d)). Each of the plots also show the Critical scale ($ S_c $) and Convergence scale ($ S_a $) along with the $ \% $ of data sampled as the sparse representation ($ D^{s}_{sparse} $) at each scale

Download: Full-size image PowerPoint slide

Figure 3. Convergence measured as the decay of upper bound (equation (45)) to the inner product in the native Hilbert space between approximation $ A_sf $ and the approximation error $ E^s $ for considered univariate ((a): TF1 and (b): TF2) and multivariate ((c): TF3 and (d): TF4) test functions

Download: Full-size image PowerPoint slide

Figure 5. Scalewise performance of Algorithm 1 on TF3. Left column ((a), (c), (e) and (g)) shows the distribution of the sparse representation selected at multiple scales. Here we have shown the projection of the sparse representation (star markers) on the X-Y plain for ease of presentation. The right column ((b), (d), (f) and (h)) shows the corresponding reconstruction for the dropwave test function from the respective $ D^s_{sparse} $. Again the fraction in the header of plots (right column) shows the proportion of dataset chosen as $ D^s_{sparse} $

Download: Full-size image PowerPoint slide

Figure 6. Confidence and Prediction Intervals for reconstruction from $ D^s_{sparse} $ from scale 0 to scale 5 for TF2. Along with the approximation produced at these scales (solid curve), the plots also show the sparse representation selected (star marked points) with the $ 95\% $ t-confidence interval (thinner dark shaded region) and $ 95\% $t-prediction interval (broader light shaded region)

Download: Full-size image PowerPoint slide

Figure 7. (Ⅰ): Importance ranking for the sparse representation (important points) sampled at scale 0 for TF2 based on ordering governed by pivoted QR decomposition; (Ⅱ) Histogram of the top 3 most important points selected for TF2 (shown as a solid curve in each subplot). Here along the columns we have the increment in scale (abbreviated as S) and along the rows we have shown the histogram of the first, second and third most important point respectively (represented with abbreviation I for Importance)

Download: Full-size image PowerPoint slide

Figure 8. Comparative behavior of the decay of $ 2 $-$ norm $ error for the Multiscale extension algorithm from [5] (represented as 'Berm') and Algorithm 1 (represented as 'Shek') for the four test function TF1, TF2, TF3 and TF4 in (a), (b), (c) and (d) respectively

Download: Full-size image PowerPoint slide

Figure 9. Performance on unseen samples for TF1, TF2, TF3 and TF4 by Algorithm 1 ($ Shek $). Here plots (a), (c), (e) and (g) ((e) and (g) just show the projection on XY plane) show the distribution of the training and testing data (50/50 % split) with corresponding plots on the right hand side ((b), (d), (f) and (g)) showing the reconstructions obtained just using the training data. The header on ((b), (d), (f) and (h)) also show the corresponding testing error

Download: Full-size image PowerPoint slide

Figure 10. Performance of Algorithm 4 from [5] ($ Berm $) on TF1, TF2, TF3 and TF4 with same experimental setup (same training and testing data as (a), (c), (e) and (g) in figure 9). Here again the corresponding testing error is displayed as a header with corresponding reconstructions from training data

Download: Full-size image PowerPoint slide

Figure 11. Generalization analysis with repeated (100 times) random splitting of data into training and testing (again 50/50 split). Here we compare the mean and standard deviation of the Root Mean Square Error (RMSE) for prediction on testing data ($ D_{test} $) by Algorithm 4 from [5] ($ Berm $) and our Algorithm 1 ($ Shek $)

Download: Full-size image PowerPoint slide

Figure 12. (a): Contour plot for Digital Elevation Model (DEM) data with 60m resolution; (b): Training DEM obtained by sampling the nodes at even indices (from DEM in (a)) resulting in a 120m resolution DEM

Download: Full-size image PowerPoint slide

Figure 13. Scale dependent reconstructions for the DEM dataset in figure 12 (a). Here (a), (b) and (c) show the reconstructed contour plot for scales 0, 2 and 6 respectively. The header on these plots also show the the proportion of dataset selected as the sparse representation along with the $ \infty-norm $ of the reconstruction error

Download: Full-size image PowerPoint slide

Figure 14. Comparison of 60m DEM reconstruction by $ Shek $ (a) and $ Berm $ (b) using the 120m DEM training data shown in figure 12 (b). The header shows the RMSE for prediction of height at the testing locations (testing RMSE for $ Shek $: 2.74 and testing RMSE for $ Berm $: 10.7)

Download: Full-size image PowerPoint slide

Algorithm 1 Hierarchical approach

1: INPUT:
      Hyperparameters: $ [(TOL \geq 0, T>0, P > 1) \in \mathbb{R}^3] $
      Dataset: $ [(X \in \mathbb{R}^{n \times d}) \subset \Omega $, $ (f|_{X} \in \mathbb{R}^n)] $
      Prediction points: $ [(X_{*} \in \mathbb{R}^{n^{*} \times d}) \subset \Omega] $
2: OUTPUT:
      Convergence Scale: $ [S_a \in \mathbb{N}] $
      Sparse Representation (

$ D_{sparse}^{S_a} $

): $ [(X_{S_a} \in \mathbb{R}^{\{l_{S_a} \times d \}}),(C_{S_a} \in \mathbb{R}^{l_{S_a}})] $
Predictions: $ [P^* \in \mathbb{R}^{n^{*}}] $
3: Initialize: $ s =0 $
4: while

$ TRUE $

do
5:      Compute covariance kernel: [$ G_s $ on $ X $ with $ (\epsilon_s = T/P^s $)]
6:      Compute numerical Rank: [$ l_s = rank(G_s) $]
7:      Remove sampling Bias: [$ (W = AG_{s}) $ with $ A \in \mathbb{R}^{k \times n} $ and $ (a_{i,j} \sim N(0,1)) $]
8:      Generate permutation information: [$ WP_R = QR $]
9:      Produce bases at scale s: $ [B^{s} = (G_{s}P_R)[:,1:l_s] $, \quad \text{with}\ $ B^{s} \in \mathbb{R}^{n \times l_s}] $
10:      Subset the sparse representation in $ X_s $
11:      Compute projection coordinate: [$ C_{s} = {(B^{s})}^{\dagger} f|_X $; $ {B^s}^{\dagger} = ({B^s}^TB^s)^{-1}{B^s}^T $]
12:      Generate approximation at s: [$ (A_sf)|_{X} = B^{s}C_s $]
13:      If $ (||f|_X- (A_sf)|_{X}||_2 \leq TOL) $ : $ S_a = s $; $ Break $
14:      Update scale: [$ s=s+1 $]
15: end while
16: Compute bases for prediction at $ X_{*} $: [$ G^{*}_{S_a} $ centered at $ X_{S_a} $with $ (\epsilon_{S_a} = T/P^{S_a} $)]
17: Predict: $ P^* = G^{*}_{S_a} C_{S_a} $

| Show Table

DownLoad: CSV

Table 1. Comparison of training and prediction time for Algorithm 1 ($ Shek $) and Algorithm 4 from [5] ($ Berm $). For training analysis, we have considered the case of different training data sizes ($ |X| $) for all test functions. Correspondingly, for testing prediction latency, we again have sets $ |X_{pred}| $ of different cardinality. For each case, the approach taking shorter time has been boldfaced

	Training time (s)			Prediction time (s)
	$ \|X\| $	$ Shek $	$ Berm $	$ \|X_{pred}\| $	$ Shek $	$ Berm $
$ TF1 $	50	4.61e-03	3.25e-02	100	1.47e-02	7.22e-02
	100	9.25e-03	1.40e-02	200	4.15e-02	3.09e-01
	150	1.60e-02	2.93e-02	300	6.21e-02	5.82e-01
	200	2.55e-02	5.63e-02	400	8.18e-02	1.08
	250	3.77e-02	9.29e-02	500	1.12e-01	1.81
	300	5.64e-02	1.25e-01	600	1.31e-01	2.38
	350	6.51e-02	1.93e-01	700	1.48e-01	3.56
	400	8.36e-02	2.21e-01	800	1.70e-01	4.15
$ TF2 $	50	5.68e-03	9.21e-03	100	2.17e-02	8.35e-02
	100	1.35e-02	1.49e-02	200	6.75e-02	3.00e-01
	150	2.86e-02	2.80e-02	300	1.53e-01	6.36e-01
	200	5.28e-02	5.27e-02	400	2.71e-01	1.07
	250	8.97e-02	8.94e-02	500	4.09e-01	1.81
	300	1.15e-01	1.23e-01	600	5.91e-01	2.29
	350	1.86e-01	1.75e-01	700	8.06e-01	3.52
	400	2.12e-01	2.17e-01	800	1.08	4.10
$ TF3 $	900	8.39e-01	7.98e-01	1225	3.13	8.67
	1600	4.19	6.40	2025	5.29	2.14e+01
	2500	2.76e+01	2.67e+01	3025	7.91	3.47e+01
	3600	7.19e+01	8.24e+01	4225	1.16e+01	4.95e+01
	4900	1.72e+02	1.99e+02	5625	1.59e+01	6.87e+01
	6400	3.84e+02	4.44e+02	7225	2.05e+01	8.73e+01
	8100	7.71e+02	8.88e+02	9025	2.75e+01	1.08e+02
	10000	1.44e+03	1.88e+03	11025	3.27e+01	2.32e+02
$ TF4 $	900	1.42	8.70e-01	1225	4.10	1.53e+01
	1600	6.94	8.68	2025	1.05e+01	3.21e+01
	2500	3.51e+01	3.44e+01	3025	2.25e+01	5.44e+01
	3600	1.27e+02	1.27e+02	4225	5.25e+01	1.32e+02
	4900	2.94e+02	2.95e+02	5625	8.78e+01	2.02e+02
	6400	7.90e+02	7.94e+02	7225	1.52e+02	3.96e+02
	8100	1.49e+03	1.49e+03	9025	2.24e+02	5.37e+02
	10000	2.69e+03	2.69e+03	11025	3.37e+02	6.90e+02

| Show Table

DownLoad: CSV

Related Papers

Cited by

References

[1]	N. I. Achieser, Theory of Approximation, Frederick Ungar Publishing Co., New York, 1956.
[2]	W. K. Allard, G. Chen and M. Maggioni, Multi-scale geometric methods for data setsⅡ: Geometric multi-resolution analysis, Appl. Comput. Harmon. Anal., 32 (2012), 435-462. doi: 10.1016/j.acha.2011.08.001.
[3]	N. Aronszajn, Theory of reproducing kernels, Trans. Amer. Math. Soc., 68 (1950), 337-404. doi: 10.1090/S0002-9947-1950-0051437-7.
[4]	F. Bellocchio, N. Borghese, S. Ferrari and V. Piuri, Kernel regression in HRBF networks for surface reconstruction, 2008 IEEE International Workshop on Haptic Audio Visual Environments and Games, (2008), 160–165. doi: 10.1109/HAVE.2008.4685317.
[5]	A. Bermanis, A. Averbuch and R. R. Coifman, Multiscale data sampling and function extension, Appl. Comput. Harmon. Anal., 34 (2013), 15-29. doi: 10.1016/j.acha.2012.03.002.
[6]	A. Bermanis, G. Wolf and A. Averbuch, Diffusion-based kernel methods on euclidean metric measure spaces, Appl. Comput. Harmon. Anal., 41 (2016), 190-213. doi: 10.1016/j.acha.2015.07.005.
[7]	B. Bohn, J. Garcke and M. Griebel, A sparse grid based method for generative dimensionality reduction of high-dimensional data, J. Comput. Phys., 309 (2016), 1-17. doi: 10.1016/j.jcp.2015.12.033.
[8]	N. A. Borghese and S. Ferrari, Hierarchical rbf networks and local parameters estimate, Neurocomputing, 19 (1998), 259-283.
[9]	C. Boutsidis, M. W. Mahoney and P. Drineas, An improved approximation algorithm for the column subset selection problem, Proceedings of the Twentieth Annual ACM-SIAM Symposium on Discrete Algorithms, SIAM, (2009), 968–977.
[10]	W. L. Briggs, V. E. Henson and S. F. McCormick, A Multigrid Tutorial, 2^nd edition, SIAM, 2000. doi: 10.1137/1.9780898719505.
[11]	M. D. Buhmann, Radial Basis Functions: Theory and Implementations, Vol. 12, Cambridge university press, 2003. doi: 10.1017/CBO9780511543241.
[12]	J. C. Carr, W. R. Fright and R. K. Beatson, Surface interpolation with radial basis functions for medical imaging, IEEE Transactions on Medical Imaging, 16 (1997), 96-107. doi: 10.1109/42.552059.
[13]	G. Chen, A. V. Little and M. Maggioni, Multi-resolution geometric analysis for data in high dimensions, in Excursions in Harmonic Analysis, Springer, 1 (2013), 259–285. doi: 10.1007/978-0-8176-8376-4_13.
[14]	E. W. Cheney, Introduction to Approximation Theory, McGraw-Hill Book Co., New York-Toronto, Ont.-London, 1966.
[15]	W. Cheney and W. Light, A Course in Approximation Theory, Vol. 101, American Mathematical Society, 2009. doi: 10.1090/gsm/101.
[16]	C. H. Chou, B. H. Kuo and F. Chang, The generalized condensed nearest neighbor rule as a data reduction method, 18th International Conference on Pattern Recognition, 2 (2006), 556-559.
[17]	A. Çivril and M. Magdon-Ismail, On selecting a maximum volume sub-matrix of a matrix and related problems, Theoret. Comput. Sci., 410 (2009), 4801-4811. doi: 10.1016/j.tcs.2009.06.018.
[18]	R. R. Coifman, S. Lafon, A. B. Lee, M. Maggioni, B. Nadler, F. Warner and S. W. Zucker, Geometric diffusions as a tool for harmonic analysis and structure definition of data: Diffusion maps, Proceedings of the National Academy of Sciences, 102 (2005), 7426-7431.
[19]	R. R. Coifman, S. Lafon, A. B. Lee, M. Maggioni, B. Nadler, F. Warner and S. W. Zucker, Geometric diffusions as a tool for harmonic analysis and structure definition of data: Multiscale methods, Proceedings of the National Academy of Sciences, 102 (2005), 7432-7437.
[20]	R. R. Coifman and M. Maggioni, Diffusion wavelets, Appl. Comput. Harmon. Anal., 21 (2006), 53-94. doi: 10.1016/j.acha.2006.04.004.
[21]	I. Daubechies, Ten Lectures on Wavelets, CBMS-NSF Regional Conference Series in Applied Mathematics, 61. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1992. doi: 10.1137/1.9781611970104.
[22]	S. De Marchi and R. Schaback, Stability of kernel-based interpolation, Adv. Comput. Math., 32 (2010), 155-161. doi: 10.1007/s10444-008-9093-4.
[23]	A. Deshpande and S. Vempala, Adaptive sampling and fast low-rank matrix approximation, in Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, Springer, (2006), 292–303. doi: 10.1007/11830924_28.
[24]	G. E. Fasshauer and J. G. Zhang, Preconditioning of radial basis function interpolation systems via accelerated iterated approximate moving least squares approximation, in Progress on Meshless Methods, Springer, (2009), 57–75. doi: 10.1007/978-1-4020-8821-6_4.
[25]	S. Ferrari, F. Bellocchio, V. Piuri and N. A. Borghese, A hierarchical rbf online learning algorithm for real-time 3-d scanner, IEEE Transactions on Neural Networks, 21 (2009), 275-285.
[26]	S. Ferrari, M. Maggioni and N. A. Borghese, Multiscale approximation with hierarchical radial basis functions networks, IEEE Transactions on Neural Networks, 15 (2004), 178-188. doi: 10.1109/TNN.2003.811355.
[27]	M. S Floater and A. Iske, Multistep scattered data interpolation using compactly supported radial basis functions, J. Comput. Appl. Math., 73 (1996), 65-78. doi: 10.1016/0377-0427(96)00035-0.
[28]	T. E. Fricker, J. E. Oakley and N. M. Urban, Multivariate Gaussian process emulators with nonseparable covariance structures, Technometrics, 55 (2013), 47-56. doi: 10.1080/00401706.2012.715835.
[29]	M. Galun, R. Basri and I. Yavneh, Review of methods inspired by algebraic-multigrid for data and image analysis applications, Numer. Math. Theory Methods Appl., 8 (2015), 283-312. doi: 10.4208/nmtma.2015.w14si.
[30]	M. Gavish, B. Nadler and R. R. Coifman, Multiscale wavelets on trees, graphs and high dimensional data: Theory and applications to semi supervised learning, ICML, (2010), 367–374.
[31]	M. Griebel and A. Hullmann, A sparse grid based generative topographic mapping for the dimensionality reduction of high-dimensional data, Modeling, Simulation and Optimization of Complex Processes-HPSC, Springer, (2012), 51–62.
[32]	M. Gu and J. O. Berger, Parallel partial Gaussian process emulation for computer models with massive output, Ann. Appl. Stat., 10 (2016), 1317-1347. doi: 10.1214/16-AOAS934.
[33]	A. Iske, Scattered data approximation by positive definite kernel functions, Rend. Semin. Mat. Univ. Politec. Torino, 69 (2011), 217-246.
[34]	P. Koumoutsakos, Multiscale flow simulations using particles, Annu. Rev. Fluid Mech., 37 (2005), 457-487. doi: 10.1146/annurev.fluid.37.061903.175753.
[35]	E. Kreyszig, Introductory Functional Analysis with Applications, Vol. 1, John Wiley & Sons, New York-London-Sydney, 1978.
[36]	D. Kushnir, M. Galun and A. Brandt, Efficient multilevel eigensolvers with applications to data analysis tasks, IEEE Transactions on Pattern Analysis and Machine Intelligence, 32 (2009), 1377-1391. doi: 10.1109/TPAMI.2009.147.
[37]	T. Lane and C. E. Brodley, Temporal sequence learning and data reduction for anomaly detection, CCS '98: Proceedings of the 5th ACM Conference on Computer and communications Security, (1998), 150–158. doi: 10.1145/288090.288122.
[38]	M. Maggioni, J. C. Bremer Jr, R. R. Coifman and A. D. Szlam, Biorthogonal diffusion wavelets for multiscale representation on manifolds and graphs, in Wavelets XI, International Society for Optics and Photonics, 5914, (2005), 59141M. doi: 10.1117/12.616909.
[39]	S. G. Mallat, A theory for multiresolution signal decomposition: The wavelet representation, IEEE Transactions on Pattern Analysis & Machine Intelligence, 11 (1989), 674-693.
[40]	S. Paul, M. Magdon-Ismail and P. Drineas, Column selection via adaptive sampling, in Advances in Neural Information Processing Systems, (2015), 406–414.
[41]	M. J. D. Powell, Approximation Theory and Methods, Cambridge university press, 1981.
[42]	A. Quarteroni and A. Veneziani, Analysis of a geometrical multiscale model based on the coupling of ODEs and PDEs for blood flow simulations, Multiscale Model. Simul., 1 (2003), 173-195. doi: 10.1137/S1540345902408482.
[43]	C. E. Rasmussen and C. K. I Williams, Gaussian Process for Machine Learning Adaptive Computation and Machine Learning, MIT Press, Cambridge, MA, 2006.
[44]	E. Sadrfaridpour, S. Jeereddy, K. Kennedy, A. Luckow, T. Razzaghi and I. Safro, Algebraic multigrid support vector machines, preprint, (2016), arXiv: 1611.05487.
[45]	L. Shih, J. D. Rennie, Y. H. Chang and D. R. Karger, Text bundling: Statistics based data-reduction, in Proceedings of the 20th International Conference on Machine Learning (ICML-03), (2003), 696–703.
[46]	K. Stüben, A review of algebraic multigrid, in Numerical Analysis: Historical Developments in the 20th Century, Elsevier, (2001), 331–359.
[47]	S. Surjanovic and D. Bingham, Virtual Library of Simulation Experiments: Test Functions and Datasets, Retrieved January 23, 2019, from http://www.sfu.ca/ ssurjano.
[48]	A. D. Szlam, M. Maggioni, R. R. Coifman and J. C. Bremer Jr, Diffusion-driven multiscale analysis on manifolds and graphs: top-down and bottom-up constructions, in Wavelets XI, International Society for Optics and Photonics, 5914 (2005), 59141D.
[49]	J. T. Vogelstein, Y. Park, T. Ohyama, R. A. Kerr, J. W. Truman, C. E. Priebe and M. Zlatic, Discovery of brainwide neural-behavioral maps via multiscale unsupervised structure learning, Science, 344 (2014), 386-392.
[50]	H. Wendland, Scattered Data Approximation, Cambridge Monographs on Applied and Computational Mathematics, Vol. 17, Cambridge university press, 2005.
[51]	Q. Wu, T. Xia, C. Chen, H.-Y. S. Lin, H. Wang and Y. Yu, Hierarchical tensor approximation of multi-dimensional visual data, IEEE Transactions on Visualization and Computer Graphics, 14 (2007), 186-199.