What you can find here:

- My papers (with available code; see also my Software and databases page)
- My lecture notes
- Theses from my students

This work has been funded in part by NSF, CITRIS, Google, Intel, Xilinx and NVIDIA Corporation. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

When available, implementations of the code (generally Matlab) and data are linked next to the corresponding paper below. If not available (particularly when it says "coming soon"), feel free to ask us and we may be able to give you some working code, even if not well documented.

Carreira-Perpiñán, M. Á. and Alizadeh, M. (2019):

"ParMAC: distributed optimisation of nested functions, with application to learning binary autoencoders".

*Conference on Systems and Machine Learning (SysML 2019)*, to appear.

[external link] [paper preprint] [slides] [poster] [C/C++/MPI implementation]Longer version: May 30, 2016, arXiv:1605.09114

[external link] [paper preprint] [slides] [C/C++/MPI implementation]Extended abstract at the Bay Area Machine Learning Symposium, Oct. 22, 2015 (BayLearn 2015):

"Distributed optimization of binary autoencoders using auxiliary coordinates"

[external link] [paper preprint] [poster]

This proposes a parallel, distributed optimisation model for the method of auxiliary coordinates (MAC) and implements it in MPI to learn binary autoencoders. We give theoretical conditions for the parallel speedup to be nearly perfect and verify them experimentally in a computer cluster.

Carreira-Perpiñán, M. Á. and Tavallali, P. (2018):

"Alternating optimization of decision trees, with application to learning sparse oblique trees".

*Advances in Neural Information Processing Systems 31 (NEURIPS 2018)*, pp. 1217-1227.

[external link] [paper preprint] [supplementary material] [poster] [Matlab implementation (coming soon)]Carreira-Perpiñán, M. Á. and Idelbayev, Y. (2018):

"Learning-compression" algorithms for neural net pruning.

*IEEE Conf. Computer Vision and Pattern Recognition (CVPR 2018)*, pp. 8532-8541.

[external link] [paper preprint] [poster] [supplementary material] [Python implementation] [© IEEE]Longer version: , arXiv:

[external link] [paper preprint]Winkler, D. A., Carreira-Perpiñán, M. Á. and Cerpa, A. E. (2018):

"Plug-and-play irrigation control at scale".

*17th Int. Conf. Information Processing in Sensor Networks (IPSN 2018)*, pp. 1-12.

[external link] [paper preprint] [slides]Carreira-Perpiñán, M. Á. and Idelbayev, Y. (2017):

"Model compression as constrained optimization, with application to neural nets. Part II: quantization".

Unpublished manuscript, Jul. 13, 2017, arXiv:1707.04319.

[external link] [paper preprint]Short version at the Workshop on Optimization for Machine Learning (NIPS 2017)

[external link] [paper preprint] [poster]Short version at the Workshop on Machine Learning on the Phone and other Consumer Devices (NIPS 2017)

[external link] [paper preprint] [poster]Extended abstract at the Bay Area Machine Learning Symposium, Oct. 19, 2017 (BayLearn 2017)

[external link] [paper preprint] [slides] [video]

Carreira-Perpiñán, M. Á. (2017):

"Model compression as constrained optimization, with application to neural nets. Part I: general framework".

Unpublished manuscript, Jul. 5, 2017, arXiv:1707.01209.

[external link] [paper preprint]

This approaches the problem of model compression from an optimization point of view and introduces 1) a generic*model compression as constrained optimization*framework, and 2) a*learning-compression (LC)*algorithm to solve it that is simple, efficient and convergent to a local optimum (under some assumptions). It allows one to build model compression algorithms for a variety of combinations of loss function, model and compression technique, by simply reusing existing code from signal compression and machine learning as a black box.Raziperchikolaei, R. and Carreira-Perpiñán, M. Á. (2017):

"Learning supervised binary hashing: optimization vs diversity".

*IEEE Int. Conf. Image Processing (ICIP 2017)*, pp. 3695-3699.

[external link] [paper preprint] [poster]Raziperchikolaei, R. and Carreira-Perpiñán, M. Á. (2017):

"Learning circulant support vector machines for fast image search".

*IEEE Int. Conf. Image Processing (ICIP 2017)*, pp. 385-389.

[external link] [paper preprint] [slides] [Matlab implementation (coming soon)]Short version at the Workshop on Optimization for Machine Learning (NIPS 2017)

[external link] [paper preprint] [poster]

Vladymyrov, M. and Carreira-Perpiñán, M. Á. (2017):

"Fast, accurate spectral clustering using locally linear landmarks".

*Int. Joint Conf. Neural Networks (IJCNN 2017)*, pp. 3870-3879.

[external link] [paper preprint]Raziperchikolaei, R. and Carreira-Perpiñán, M. Á. (2016):

"Learning independent, diverse binary hash functions: pruning and locality".

*17th IEEE Int. Conf. Data Mining (ICDM 2016)*, pp. 1173-1178.

[external link] [paper preprint] [slides] [Matlab implementation]Carreira-Perpiñán, M. Á. and Raziperchikolaei, R. (2016):

"An ensemble diversity approach to supervised binary hashing".

*Advances in Neural Information Processing Systems 29 (NIPS 2016)*, pp. 757-765.

[external link] [paper preprint] [supplementary material] [poster] [Matlab implementation]Longer version: Feb. 3, 2016, arXiv:1602.01557

[external link] [paper preprint]Extended abstract at the Bay Area Machine Learning Symposium, Oct. 22, 2015 (BayLearn 2015):

[external link] [paper preprint] [slides] [video]

This shows that competitive binary hash functions (whose output is a vector of binary values) can be learned

*independently*by simply making them differ through ensemble diversity techniques. This is far faster and scalable than solving a joint optimisation problem where all bits are coupled.Raziperchikolaei, R. and Carreira-Perpiñán, M. Á. (2016):

"Optimizing affinity-based binary hashing using auxiliary coordinates".

*Advances in Neural Information Processing Systems 29 (NIPS 2016)*, pp. 640-648.

[external link] [paper preprint] [supplementary material] [poster] [Matlab implementation]Longer version: Feb. 5, 2016, arXiv:1501.05352

[external link] [paper preprint]Short version at the Workshop on Non-Convex Analysis and Optimization (ICML 2016):

[external link] [paper preprint] [poster]Extended abstract at the Bay Area Machine Learning Symposium, Oct. 6, 2016 (BayLearn 2016):

[external link] [paper preprint] [poster]

This uses the method of auxiliary coordinates (MAC) to learn an arbitrary binary hash function (i.e., a function whose output is a vector of binary values) with an arbitrary affinity-based loss function.

Vladymyrov, M. and Carreira-Perpiñán, M. Á. (2016):

"The Variational Nyström method for large-scale spectral problems".

*33rd Int. Conf. Machine Learning (ICML 2016)*, pp. 211-220.

[external link] [paper preprint] [supplementary material] [slides] [poster] [Matlab implementation]Winkler, D. A., Wang, R., Blanchette, F., Carreira-Perpiñán, M. Á. and Cerpa, A. E. (2016):

"MAGIC: Model-Based Actuation for Ground Irrigation Control".

*15th Int. Conf. Information Processing in Sensor Networks (IPSN 2016)*, article No. 9.

[external link] [paper preprint] [slides]Poster abstract at the

*13th ACM Conf. Embedded Networked Sensor Systems (SenSys 2015)*:

"MICO: Model-Based Irrigation Control Optimization", pp. 409-410

[external link] [paper preprint]

Carreira-Perpiñán, M. Á. and M. Vladymyrov (2015):

"A fast, universal algorithm to learn parametric nonlinear embeddings".

*Advances in Neural Information Processing Systems 28 (NIPS 2015)*, pp. 253-261.

[external link] [paper preprint] [supplementary material] [poster] [Matlab implementation (coming soon)]Extended abstract at the Bay Area Machine Learning Symposium, Oct. 21, 2014 (BayLearn 2014):

[external link] [paper preprint] [poster]

This uses the method of auxiliary coordinates (MAC) to learn an optimal mapping (such as linear or a neural net) for a nonlinear embedding (such as the elastic embedding or t-SNE).

Carreira-Perpiñán, M. Á. (2015):

"Clustering methods based on kernel density estimators: mean-shift algorithms".

Invited chapter in*Handbook of Cluster Analysis*(C. Hennig, M. Meila, F. Murtagh and R. Rocci, eds.), CRC/Chapman and Hall, chapter 18, pp. 383-418.

[external link] [paper preprint]Preprint version:

Carreira-Perpiñán, M. Á. (2015):

"A review of mean-shift algorithms for clustering".

Unpublished manuscript, Mar. 2, 2015, arXiv:1503.00687.

[external link] [paper preprint]

This contains a review of mean-shift algorithms, mostly for nonparametric clustering but also for problems such as manifold denoising and multivalued regression (modal regression). It also discusses relations with scale-space theory, spectral clustering and

*K*-modes algorithms, as well as theoretical results about mean-shift algorithms and the geometry of modes of kernel density estimators and Gaussian mixtures.Archambault, D., Bunte, K. Carreira-Perpiñán, M. Á., Ebert, D., Ertl, T. and Zupan, B. (2015):

"Machine learning meets visualization: a roadmap for scalable data analytics".

In*Bridging Information Visualization with Machine Learning*(D. A. Keim, T. Munzner, F. Rossi and M. Verleysen, eds.),*Dagstuhl Reports***5**(3):8-12.

[external link] [paper preprint]Carreira-Perpiñán, M. Á. and Raziperchikolaei, R. (2015):

"Hashing with binary autoencoders".

*IEEE Conf. Computer Vision and Pattern Recognition (CVPR 2015)*, pp. 557-566.

[external link] [paper preprint] [slides] [poster] [Matlab implementation] [Weka implementation]Longer version: Jan. 5, 2015, arXiv:1501.00756

[external link] [paper preprint] [slides]Extended abstract at the Bay Area Machine Learning Symposium, Oct. 21, 2014 (BayLearn 2014):

[external link] [paper preprint] [slides] [poster]Workshop paper at the 2015 INFORMS Workshop on Data Mining and Analytics:

[external link] [paper preprint] [slides]Short version at the Workshop on Optimization Methods for the Next Generation of Machine Learning (ICML 2016):

"Optimizing binary autoencoders using auxiliary coordinates, with application to learning binary hashing"

[external link] [paper preprint] [poster]

This uses the method of auxiliary coordinates (MAC) to learn an autoencoder with binary hidden units, whose encoder gives good binary hash functions for image retrieval.

Carreira-Perpiñán, M. Á. (2014):

"An ADMM algorithm for solving a proximal bound-constrained quadratic program".

Unpublished manuscript, Dec. 29, 2014, arXiv:1412.8493 [math.OC].

[external link] [paper preprint] [Matlab implementation]Wang, W. and Carreira-Perpiñán, M. Á. (2014):

"The Laplacian*K*-modes algorithm for clustering".

Unpublished manuscript, Jun. 15, 2014, arXiv:1406.3895.

[external link] [paper preprint] [slides] [animations] [Matlab implementation]Wang, W. and Carreira-Perpiñán, M. Á. (2014):

"The role of dimensionality reduction in classification".

*28th AAAI Conference on Artificial Intelligence (AAAI 2014)*, pp. 2128-2134.

[external link] [paper preprint] [slides] [poster] [animations] [Matlab implementation] [Weka implementation]Longer version:

Wang, W. and Carreira-Perpiñán, M. Á.:

"The role of dimensionality reduction in classification" (2014).

Unpublished manuscript, May 25, 2014, arXiv:1405.6444.

[external link] [paper preprint]

This uses the method of auxiliary coordinates (MAC) to learn nonlinear, low-dimensional features for a linear SVM.

Carreira-Perpiñán, M. Á. and Wang, W. (2014):

"LASS: a simple assignment model with Laplacian smoothing".

*28th AAAI Conference on Artificial Intelligence (AAAI 2014)*, pp. 1715-1721.

[external link] [paper preprint] [slides] [poster] [Matlab implementation]Longer version:

Carreira-Perpiñán, M. Á. and Wang, W.:

"LASS: a simple assignment model with Laplacian smoothing" (2014).

Unpublished manuscript, May 22, 2014, arXiv:1405.5960.

[external link] [paper preprint]Extended abstract at the Bay Area Machine Learning Symposium, Aug. 28, 2013 (BayLearn 2013):

"A simple assignment model with Laplacian smoothing"

[external link] [paper preprint] [poster]

LASS learns soft assignments of

*N*items to*K*categories, given sparse similarity information between items, and between items and categories (as e.g. in image or document tagging). It can be seen as semi-supervised learning of assignments or probability distributions.Carreira-Perpiñán, M. Á. and Wang, W. (2014):

"Distributed optimization of deeply nested systems".

*17th Int. Conf. Artificial Intelligence and Statistics (AISTATS 2014)*, pp. 10-19.

**Notable paper award**(even more interesting, this paper was rejected twice from NIPS and twice from ICML).

[external link] [paper preprint] [supplementary material] [slides] [video] [poster] [Matlab implementation (coming soon)]Longer version:

Carreira-Perpiñán, M. Á. and Wang, W.:

"Distributed optimization of deeply nested systems" (2012).

Unpublished manuscript, Dec. 24, 2012, arXiv:1212.5921.

[external link] [paper preprint]Extended abstract at the Bay Area Machine Learning Symposium, Aug. 30, 2012 (BayLearn 2012):

"Fast algorithms for learning deep neural networks"

[external link] [paper preprint] [slides] [poster]

This introduces a new, general mathematical technique, the

*method of auxiliary coordinates (MAC)*, to train deep nets and other "nested" systems. It results in a*coordination-minimisation (CM) algorithm*that alternates training individual layers (reusing existing algorithms) with glueing layers (using auxiliary coordinates). It is provably convergent, simple to implement, has embarrassing parallelism within each step, and can apply even if derivatives are not available. It can also do model selection "on the fly" and search over architectures while it trains.Vladymyrov, M. and Carreira-Perpiñán, M. Á. (2014):

"Linear-time training of nonlinear low-dimensional embeddings".

*17th Int. Conf. Artificial Intelligence and Statistics (AISTATS 2014)*, pp. 968-977.

[external link] [paper preprint] [supplementary material] [slides] [poster] [Matlab implementation]Extended abstract at the Bay Area Machine Learning Symposium, Aug. 28, 2013 (BayLearn 2013):

"Linear-time training of nonlinear low-dimensional embeddings"

[external link] [paper preprint] [poster]

This uses fast multipole methods to approximate the gradient computation in linear time, so we can train embeddings with millions of points in a few hours.

Erickson, V., Carreira-Perpiñán, M. Á. and Cerpa, A. E. (2014):

"Occupancy modeling and prediction for building energy management".

*ACM Trans. Sensor Networks*,**10**(3):42.1-42.28.

[external link] [paper preprint]Wang, W. and Carreira-Perpiñán, M. Á. (2013):

"Projection onto the probability simplex: an efficient algorithm with a simple proof, and an application".

Unpublished manuscript, Sep. 3, 2013, arXiv:1309.1541.

[external link] [paper preprint] [Matlab implementation]Vladymyrov, M. and Carreira-Perpiñán, M. Á. (2013):

"Locally Linear Landmarks for large-scale manifold learning".

*24th European. Conf. Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD 2013)*, pp. 256-271.

[external link] [paper preprint] [slides] [poster] [Matlab implementation]Short version at the Workshop on Spectral Learning (ICML 2013):

[external link] [paper preprint] [poster]

LLL is a fast, approximate method to solve spectral dimensionality reduction problems such as Laplacian eigenmaps. It is faster than the Nyström formula while giving more accurate results.

Vladymyrov, M. and Carreira-Perpiñán, M. Á. (2013):

"Entropic affinities: properties and efficient numerical computation".

*30th Int. Conf. Machine Learning (ICML 2013)*, pp. 477-485.

[external link] [paper preprint] [supplementary material] [slides] [video] [poster] [Matlab implementation]

The*entropic affinities*(introduced by Hinton and Roweis, NIPS 2002) are a way to construct Gaussian affinities with an adaptive bandwidth for each data point, so that each point has a fixed effective number of neighbours (perplexity)*K*. As we show, they can be computed efficiently. They work better than using a global bandwidth for all points in problems such as nonlinear embeddings, spectral clustering, etc.Kamthe, A., Carreira-Perpiñán, M. Á. and Cerpa, A. E. (2013):

"Improving wireless link simulation using multi-level Markov models".

*ACM Trans. Sensor Networks***10**(1):17.1-17.28.

[external link] [paper preprint] [TOSSIM code for TinyOS 2.0]Carreira-Perpiñán, M. Á. and Wang, W. (2013):

"The*K*-modes algorithm for clustering".

Unpublished manuscript, Apr. 23, 2013, arXiv:1304.6478.

[external link] [paper preprint] [slides] [Matlab implementation]Kamthe, A., Carreira-Perpiñán, M. Á. and Cerpa, A. E. (2013):

"Quick construction of data-driven models of the short-term behavior of wireless links".

*32nd Annual IEEE Int. Conf. Computer Communications (INFOCOM 2013) Mini-Conference*, 160-164.

[external link] [paper preprint] [slides]Vladymyrov, M. and Carreira-Perpiñán, M. Á. (2012):

"Partial-Hessian strategies for fast learning of nonlinear embeddings".

*29th Int. Conf. Machine Learning (ICML 2012)*, pp. 345-352.

[external link] [paper preprint] [supplementary material] [slides] [video] [poster] [Matlab implementation] [R implementation by James Melville]Extended abstract at the Bay Area Machine Learning Symposium, Aug. 30, 2012 (BayLearn 2012):

"Fast training of graph-based algorithms for nonlinear dimensionality reduction"

[external link] [paper preprint] [slides]

This introduces the

*spectral direction*, a fast optimisation method for nonlinear embedding methods. This code is the fastest one available for EE, SNE and t-SNE as far as I know.Wang, W. and Carreira-Perpiñán, M. Á. (2012):

"Nonlinear low-dimensional regression using auxiliary coordinates".

*15th Int. Conf. Artificial Intelligence and Statistics (AISTATS 2012)*, pp. 1295-1304.

[external link] [paper preprint] [poster] [animations] [Matlab implementation (coming soon)]Farhadloo, M. and Carreira-Perpiñán, M. Á. (2012):

"Learning and adaptation of a tongue shape model with missing data".

*IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP 2012)*, pp. 3981-3984.

[external link] [paper preprint] [slides] [Matlab implementation (coming soon)] [© IEEE]Farhadloo, M. and Carreira-Perpiñán, M. Á. (2012):

"Regularising an adaptation algorithm for tongue shape models".

*IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP 2012)*, pp. 4481-4484.

[external link] [paper preprint] [poster] [Matlab implementation (coming soon)] [© IEEE]Wang, W., Carreira-Perpiñán, M. Á. and Lu, Z. (2011):

"A denoising view of matrix completion".

*Advances in Neural Information Processing Systems 24 (NIPS 2011)*, pp. 334-342.

[external link] [paper preprint] [poster] [animations] [Matlab implementation]Carreira-Perpiñán, M. Á. and Lu, Z. (2011):

"Manifold learning and missing data recovery through unsupervised regression".

*12th IEEE Int. Conf. Data Mining (ICDM 2011)*, pp. 1014-1019.

[external link] [paper preprint] [slides] [animations] [Matlab implementation] [© IEEE]Kamthe, A., Erickson, V., Carreira-Perpiñán, M. Á. and Cerpa, A. E. (2011):

"Enabling building energy auditing using adapted occupancy models".

*3rd ACM Workshop on Embedded Sensing Systems for Energy-Efficiency in Buildings (BuildSys 2011)*, pp. 31-36.

[external link] [paper preprint] [slides]Kamthe, A., Carreira-Perpiñán, M. Á. and Cerpa, A. E. (2011):

"Adaptation of a mixture of multivariate Bernoulli distributions".

*22nd Int. Joint Conf. Artificial Intelligence (IJCAI 2011)*, pp. 1336-1341.

[external link] [paper preprint] [poster] [Matlab implementation (coming soon)]Erickson, V., Carreira-Perpiñán, M. Á. and Cerpa, A. E. (2011):

"OBSERVE: Occupancy-Based System for Efficient Reduction of HVAC Energy".

*10th Int. Conf. Information Processing in Sensor Networks (IPSN 2011)*, pp. 258-269.

[external link] [paper preprint] [slides] [poster]Carreira-Perpiñán, M. Á. (2011):

"Machine learning models of the tongue shape during speech".

*9th Int. Seminar on Speech Production (ISSP 2011)*, pp. 103-110.

[external link.] [paper preprint] [poster]Qin, C., Carreira-Perpiñán, M. Á. and Farhadloo, M. (2010):

"Adaptation of a tongue shape model by local feature transformations".

*Interspeech 2010*, pp. 1596-1599.

[external link] [paper preprint] [poster] [Matlab implementation (coming soon)]Qin, C. and Carreira-Perpiñán, M. Á. (2010):

"Estimating missing data sequences in X-ray microbeam recordings".

*Interspeech 2010*, pp. 1592-1595.

[external link] [paper preprint] [poster] [Matlab implementation]

The preprint fixes a typo in the 3rd eq. of p. 2.Qin, C. and Carreira-Perpiñán, M. Á. (2010):

"Articulatory inversion of American English /r/ by conditional density modes".

*Interspeech 2010*, pp. 1998-2001.

[external link] [paper preprint] [slides] [animations]Xie, L., Carreira-Perpiñán, M. Á. and Newsam, S. (2010):

"Semi-supervised regression with temporal image sequences".

*IEEE Int. Conf. Image Processing (ICIP 2010)*, pp. 2637-2640.

[external link] [paper preprint] [slides] [© IEEE]Carreira-Perpiñán, M. Á. (2010):

"The elastic embedding algorithm for dimensionality reduction".

*27th Int. Conf. Machine Learning (ICML 2010)*, pp. 167-174.

[external link] [paper preprint] [supplementary material] [slides] [poster] [Matlab implementation]

Note: the code in the link above (which uses a fixed-point iteration method) has been superseded by a faster one based on the spectral direction iteration; see the newer code from the paper by Vladymyrov and Carreira-Perpiñán, ICML 2012.Wang, W. and Carreira-Perpiñán, M. Á. (2010):

"Manifold blurring mean shift algorithms for manifold denoising".

*IEEE Conf. Computer Vision and Pattern Recognition (CVPR 2010)*, pp. 1759-1766.

[external link] [paper preprint] [slides] [videolectures.net] [poster] [animations] [Matlab implementation (coming soon)] [© IEEE]Carreira-Perpiñán, M. Á. and Lu, Z. (2010):

"Parametric dimensionality reduction by unsupervised regression".

*IEEE Conf. Computer Vision and Pattern Recognition (CVPR 2010)*, pp. 1895-1902.

[external link] [paper preprint] [poster] [animations] [Matlab implementation] [© IEEE]Qin, C. and Carreira-Perpiñán, M. Á. (2010):

"Reconstructing the full tongue contour from EMA/X-ray microbeam".

*IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP 2010)*, pp. 4190-4193.

[external link] [paper preprint] [slides] [animations] [Matlab implementation (coming soon)] [© IEEE]Massaro, D. A., Carreira-Perpiñán, M. Á. and D. J. Merrill (2010):

"An automatic visible speech supplement for deaf individuals' speech comprehension in face-to-face and classroom situations".

In*Cued Speech and Cued Language Development of Deaf and Hard of Hearing Children*(C. LaSasso, K. Lamar Crain and J. Leybaert, eds.), Plural Publishing, chapter 22, pp. 503-530.

[external link] [paper preprint]Qin, C. and Carreira-Perpiñán, M. Á. (2009):

"The geometry of the articulatory region that produces a speech sound".

Invited paper,*43rd Annual Asilomar Conf. Signals, Systems, and Computers*, pp. 1742-1746.

[external link] [paper preprint] [slides] [© IEEE]Qin, C. and Carreira-Perpiñán, M. Á. (2009):

"Adaptation of a predictive model of tongue shapes".

*Interspeech 2009*, pp. 772-775.

[external link] [paper preprint] [poster]Kamthe, A., Carreira-Perpiñán, M. Á. and Cerpa, A. E. (2009):

"M&M: Multi-level Markov model for wireless link simulations".

*7th ACM Conf. Embedded Networked Sensor Systems (SenSys 2009)*, pp. 57-70.

[external link] [paper preprint] [slides] [poster] [TOSSIM code for TinyOS 2.0]Poster Abstract: "Wireless link simulations using multi-level Markov models", pp. 391-392.

[external link] [paper preprint]

Massaro, D. A., Carreira-Perpiñán, M. Á. and D. J. Merrill (2009):

"Optimizing visual feature perception for an automatic wearable speech supplement in face-to-face communication and classroom situations".

*42st Hawaii Int. Conf. Systems Science (HICSS-42 2009)*, pp. 1-10.

[external link] [paper preprint] [© IEEE]Massaro, D. A., Carreira-Perpiñán, M. Á., D. J. Merrill, C. Sterling, S. Bigler, E. Piazza and M. Perlman (2008):

"iGlasses: an automatic wearable speech supplement in face-to-face communication and classroom situations" (demonstration paper).

*Int. Conf. Multimodal Interfaces (ICMI 2008)*, pp. 197-198.

[external link] [paper preprint]Qin, C., Carreira-Perpiñán, M. Á., Richmond, K., Wrench, A. and Renals, S. (2008):

"Predicting tongue shapes from a few landmark locations".

*Interspeech 2008*, pp. 2306-2309.

[external link] [paper preprint] [slides]Carreira-Perpiñán, M. Á. and Lu, Z. (2008):

"Dimensionality reduction by unsupervised regression".

*IEEE Conf. Computer Vision and Pattern Recognition (CVPR 2008)*.

[external link] [paper preprint] [slides] [poster] [animations] [Matlab implementation] [© IEEE]

The preprint fixes three errata in the legends of fig. 2.Carreira-Perpiñán, M. Á. (2008):

"Generalised blurring mean-shift algorithms for nonparametric clustering".

*IEEE Conf. Computer Vision and Pattern Recognition (CVPR 2008)*.

[external link] [paper preprint] [poster] [animations] [© IEEE]Lu, Z. and Carreira-Perpiñán, M. Á. (2008):

"Constrained spectral clustering through affinity propagation".

*IEEE Conf. Computer Vision and Pattern Recognition (CVPR 2008)*.

[external link] [paper preprint] [poster] [Matlab implementation] [© IEEE]Qin, C. and Carreira-Perpiñán, M. Á. (2008):

"Trajectory inverse kinematics by conditional density modes".

*IEEE Int. Conf. on Robotics and Automation (ICRA 2008)*, pp. 1979-1986.

[external link] [paper preprint] [slides] [poster] [animations] [© IEEE]Qin, C. and Carreira-Perpiñán, M. Á. (2008):

"Trajectory inverse kinematics by nonlinear, nongaussian tracking".

*IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP 2008)*, pp. 2057-2060.

[external link] [paper preprint] [poster] [© IEEE]Özertem, U., Erdogmus, D. and Carreira-Perpiñán, M. Á. (2008):

"Density geodesics for similarity clustering".

*IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP 2008)*, pp. 1977-1980.

[external link] [paper preprint] [© IEEE]Lu, Z., Carreira-Perpiñán, M. Á. and Sminchisescu, C. (2008):

"People tracking with the Laplacian Eigenmaps Latent Variable Model".

*Advances in Neural Information Processing Systems 20 (NIPS 2007)*, pp. 1705-1712.

[external link] [paper preprint] [poster] [slides] [animations]

Preview at The Learning Workshop 2007: [external link] [paper preprint] [poster]Qin, C. and Carreira-Perpiñán, M. Á. (2007):

"An empirical investigation of the nonuniqueness in the acoustic-to-articulatory mapping".

*Interspeech 2007*, pp. 74-77.

**Best student paper award**.

[external link] [paper preprint] [slides]Qin, C. and Carreira-Perpiñán, M. Á. (2007):

"A comparison of acoustic features for articulatory inversion".

*Interspeech 2007*, pp. 2469-2472.

[external link] [paper preprint] [slides]Kazmierczak, S. C., Leen, T. K., Erdogmus, D. and Carreira-Perpiñán, M. Á. (2007):

"Reduction of multi-dimensional laboratory data to a two-dimensional plot: a novel technique for the identification of laboratory error".

*Clinical Chemistry and Laboratory Medicine***45**(6):749-752.

[external link] [paper preprint]Myronenko, A., Song, X. and Carreira-Perpiñán, M. Á. (2007):

"Free-form nonrigid image registration using generalized elastic nets".

*IEEE Conf. Computer Vision and Pattern Recognition (CVPR 2007)*.

[external link] [paper preprint] [poster] [© IEEE]Carreira-Perpiñán, M. Á. and Lu, Z. (2007):

"The Laplacian Eigenmaps Latent Variable Model".

*11th Int. Conf. Artificial Intelligence and Statistics (AISTATS 2007)*, pp. 59-66.

[external link] [paper preprint] [poster] [slides] [animations] [Matlab implementation (coming soon)]Myronenko, A., Song, X. and Carreira-Perpiñán, M. Á. (2007):

"Non-rigid point set registration: Coherent Point Drift".

*Advances in Neural Information Processing Systems 19 (NIPS 2006)*, pp. 1009-1016.

[external link] [paper preprint] [poster] [Matlab implementation]Carreira-Perpiñán, M. Á. (2007):

"Gaussian mean shift is an EM algorithm".

*IEEE Trans. on Pattern Analysis and Machine Intelligence***29**(5):767-776.

[external link] [paper preprint] [© IEEE]

We prove the following properties of the mean-shift algorithm: (1) Mean-shift is an expectation-maximisation (EM) algorithm for Gaussian kernels and a generalised EM algorithm for non-gaussian kernels. (2) Gaussian mean-shift converges linearly to a mode from almost any starting point; the rate of linear convergence approaches 0 (superlinear convergence) for very narrow or very wide kernels, but is often close to 1 (thus extremely slow) for intermediate widths, and exactly 1 (sublinear convergence) for widths at which modes merge. (3) The iterates approach the mode along the local principal component of the data points from the inside of the convex hull of the data points. (4) The convergence domains are nonconvex and can be disconnected and show fractal behaviour.Carreira-Perpiñán, M. Á. (2006):

"Fast nonparametric clustering with Gaussian blurring mean-shift".

*23rd Int. Conf. Machine Learning (ICML 2006)*, pp. 153-160.

[external link] [paper preprint] [slides] [poster] [animations] [Java applet] [Matlab implementation]

We give a stopping criterion and an acceleration for the Gaussian blurring mean-shift algorithm, prove it has cubic convergence rate with Gaussian clusters, connect it with spectral clustering, and show it achieves image segmentations as good as those of Gaussian mean-shift but much faster.Carreira-Perpiñán, M. Á. (2006):

"Acceleration strategies for Gaussian mean-shift image segmentation".

*IEEE Conf. Computer Vision and Pattern Recognition (CVPR 2006)*, pp. 1160-1167.

[external link] [paper preprint] [poster] [Java applet] [Matlab implementation (coming soon)] [© IEEE]

We attain speedups of one to two orders of magnitude over exact Gaussian mean-shift*while achieving essentially the same segmentation*, by using techniques based on spatial discretisation, spatial neighbourhoods, sparse EM and EM-Newton algorithms.Myronenko, A., Song, X. and Carreira-Perpiñán, M. Á. (2006):

"Non-parametric image registration using generalized elastic nets".

*9th MICCAI Conference, Int. Workshop on Mathematical Foundations of Computational Anatomy (MFCA 2006)*, pp. 156-163.

[external link] [paper preprint]Erdogmus, D., Carreira-Perpiñán, M. Á. and Özertem, U. (2006):

"Kernel density estimation, affinity-based clustering, and typical cuts".

*IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP'06)*, vol. 5 pp. 569-572.

[external link] [paper preprint] [© IEEE]Carreira-Perpiñán, M. Á., Dayan, P. and Goodhill, G. J. (2005):

"Differential priors for elastic nets".

*6th Int. Conf. Intelligent Data Engineering and Automated Learning (IDEAL'05)*, pp. 335-342, Lecture Notes in Computer Science vol. 3578, Springer-Verlag.

[external link] [paper preprint] [Matlab implementation] [supplementary information] [© Springer-Verlag]Carreira-Perpiñán, M. Á. and Hinton, G. E. (2005):

"On contrastive divergence learning".

*10th Int. Workshop on Artificial Intelligence and Statistics (AISTATS 2005)*, pp. 59-66, 2005.

[external link] [paper preprint] [poster]

Contrastive divergence (CD) is a fast, low-variance Markov chain Monte Carlo method for maximum likelihood estimation of random fields. We show that CD is typically slightly biased and give an unbiased alternative algorithm.Carreira-Perpiñán, M. Á. and Zemel, R. S. (2005):

"Proximity graphs for clustering and manifold learning".

*Advances in Neural Information Processing Systems 17 (NIPS 2004)*, pp. 225-232.

[external link] [paper preprint] [slides] [poster]

We propose new types of proximity (neighbourhood) graphs, based on ensembles of minimum spanning trees, for use with (spectral) clustering and manifold learning.Carreira-Perpiñán, M. Á., Lister, R. J., and Goodhill, G. J. (2005):

"A computational model for the development of multiple maps in primary visual cortex".

*Cerebral Cortex***15**(8):1222-1233.

[external link] [paper preprint] [Matlab implementation] [supplementary information]

We model the combined development of 5 maps of primary visual cortex (retinotopy, ocular dominance, orientation, direction and spatial frequency) using the elastic net model, as well as the effects of monocular deprivation and single-orientation rearing.*We also predict that the stripe width of all maps (orientation, direction, spatial frequency) increases slightly under monocular deprivation*. This prediction has been confirmed by Farley et al.,*J. Neurosci.*2007.Carreira-Perpiñán, M. Á. and Goodhill, G. J. (2004):

"Influence of lateral connections on the structure of cortical maps".

*J. Neurophysiology***92**(5):2947-2959.

[external link] [paper preprint] [Matlab implementation] [supplementary information]

Using a generalised elastic net model of cortical maps, we show that the number of excitatory and inhibitory oscillations of a Mexican-hat cortical interaction function has a remarkable effect on the geometric relations between the maps of ocular dominance and orientation.*We predict that, in biological maps, this function oscillates only once (central excitation, surround inhibition)*.He, X., Zemel, R. S., and Carreira-Perpiñán, M. Á. (2004):

"Multiscale conditional random fields for image labeling".

*IEEE Conf. Computer Vision and Pattern Recognition (CVPR 2004)*, pp. 695-702, Washington, DC, 27 June - 2 July 2004.

[external link] [paper preprint] [Corel and Sowerby datasets (gzipped Matlab format)] [© IEEE]Carreira-Perpiñán, M. Á.:

"Reconstruction of sequential data with density models".

Unpublished manuscript, Jan. 27, 2004, arXiv:1109.3248.

[external link] [paper preprint]Carreira-Perpiñán, M. Á. and Goodhill, G. J.:

"Generalised elastic nets".

Unpublished manuscript, Aug. 14, 2003, arXiv:1108.2840 [q-bio.NC].

[external link] [paper preprint] [Matlab implementation] [supplementary information]Carreira-Perpiñán, M. Á. and Williams, C. K. I. (2003):

"An isotropic Gaussian mixture can have more modes than components".

Technical report EDI-INF-RR-0185, School of Informatics, University of Edinburgh, UK.

[external link] [paper] [supplementary information]Carreira-Perpiñán, M. Á. and Williams, C. K. I. (2003):

"On the number of modes of a Gaussian mixture".

*Scale-Space Methods in Computer Vision*, pp. 625-640, Lecture Notes in Computer Science vol. 2695, Springer-Verlag.

[external link] [paper preprint] [© Springer-Verlag] [extended technical report version] [supplementary information]Carreira-Perpiñán, M. Á. and Williams, C. K. I. (2003):

"On the number of modes of a Gaussian mixture".

Technical report EDI-INF-RR-0159, School of Informatics, University of Edinburgh, UK.

[external link] [paper] [supplementary information]Carreira-Perpiñán, M. Á. and Goodhill, G. J. (2002):

"Are visual cortex maps optimized for coverage?".

*Neural Computation***14**(7):1545-1560.

[external link] [paper preprint]Carreira-Perpiñán, M. Á. and Goodhill, G. J. (2002):

"Development of columnar structures in primary visual cortex".

Invited chapter in*Computational Neuroanatomy: Principles and Methods*(G. A. Ascoli, ed.), Humana Press, chapter 15, pp. 337-357.

[external link] [paper preprint]

This contains a review of data and models for visual cortical maps, with an emphasis on the elastic net.Goodhill, G. J. and Carreira-Perpiñán, M. Á. (2002):

"Cortical Columns".

Invited article in the*Encyclopedia of Cognitive Science*(L. Nadel, ed.), Macmillan, vol. 1, pp. 845-851.

[external link] [paper preprint]

This contains a review of columnar systems in the cortex, in particular in the visual cortex (maps of ocular dominance, orientation, etc.).Carreira-Perpiñán, M. Á. and Goodhill, G. J. (2001):

"The effect of variable elastic topologies on the structure of ocular dominance and orientation maps".*Society for Neuroscience abstracts*,**27**, 475.21.

[external link] [poster] [supplementary information]Carreira-Perpiñán, M. Á. (2001):

"Continuous latent variable models for dimensionality reduction and sequential data reconstruction".

PhD thesis, University of Sheffield, UK.

[external link] [abstract] [paper PDF] [paper PS] [list of all references (.tar.gz, BibTeX)]

The following chapters, available separately, contain tutorial material:- Chapter 2:
*The continuous latent variable modelling formalism*.

This contains a review of continuous latent variable models: probabilistic principal component analysis (PCA), factor analysis, the generative topographic mapping (GTM), independent component analysis (ICA), mixtures of latent variable models, etc. It also deals with issues such as parameter estimation, identifiability, interpretability, visualisation, and dimensionality reduction with continuous latent variable models.

[paper PDF] [paper PS] - Chapter 4:
*Dimensionality reduction*.

This contains a review of dimensionality reduction with nonprobabilistic methods (probabilistic methods, i.e., latent variable models, are reviewed in chapter 2): nonlinear autoassociators, kernel PCA, principal curves, vector quantisation, multidimensional scaling, Isomap, LLE, etc. It also reviews issues such as the curse of dimensionality and the intrinsic dimensionality.

[paper PDF] [paper PS] - Chapter 6:
*Inverse problems and mapping inversion*.

This contains a review of inverse problem theory and its methods (Bayesian and nonprobabilistic), and its relation with latent variable models.

[paper PDF] [paper PS] - Chapter 10 (partially):
*The acoustic-to-articulatory mapping problem*.

This contains a review of the acoustic-to-articulatory mapping problem of speech research (the recovery of the acoustic waveform given the vocal tract configuration), emphasising its potential role in improving automatic speech recognition.

[paper PDF] [paper PS]

- Chapter 2:
Carreira-Perpiñán, M. Á. (2000):

"Mode-finding for mixtures of Gaussian distributions".

*IEEE Trans. on Pattern Analysis and Machine Intelligence***22**(11):1318-1323.

[external link] [paper preprint] [© IEEE] [extended technical report version] [Matlab implementation]

This gives algorithms for finding (nearly)*all*the maxima of a Gaussian mixture.Carreira-Perpiñán, M. Á. (2000):

"Reconstruction of sequential data with probabilistic models and continuity constraints".

*Advances in Neural Information Processing Systems 12 (NIPS'99)*, pp. 414-420.

[external link] [paper preprint] [slides]

This gives a method for reconstructing a sequence of data vectors where some components at some times are missing, with applications to inverse problems in speech (articulatory inversion) and robotics (inverse kinematics).Carreira-Perpiñán, M. Á. and Renals, S. (1999):

"A latent variable modelling approach to the acoustic-to-articulatory mapping problem".

*14th International Congress of Phonetic Sciences (ICPhS'99)*, pp. 2013-2016, San Francisco, USA, 1-7 August 1999.

[external link] [paper preprint] [slides]Carreira-Perpiñán, M. Á. (1999):

"One-to-many mappings, continuity constraints and latent variable models".

*1999 IEE Colloquium on Applied Statistical Pattern Recognition*, Birmingham, UK.

[external link] [paper preprint] [slides]Carreira-Perpiñán, M. Á. and Renals, S. (2000):

"Practical identifiability of finite mixtures of multivariate Bernoulli distributions".

*Neural Computation***12**(1):141-152.

[external link] [paper preprint] [Matlab implementation]

Mixtures of multivariate Bernoulli distributions are known to be nonidentifiable. We give empirical support to their "practical identifiability" and conjecture that only a small portion of the parameter space may be nonidentifiable. We also give practical advice in estimating the best number of components with an EM algorithm. This conjecture has been recently proven (Allman et al.,*Ann. Statist.*2009, Elmore et al.,*Ann. Inst. Fourier*2005): the region of nonidentifiability in parameter space has measure zero for sufficiently many variables.Carreira-Perpiñán, M. Á. (1999):

"Mode-finding for mixtures of Gaussian distributions" (revised August 4, 2000).

Technical report CS-99-03, Dept. of Computer Science, University of Sheffield, UK.

[external link] [paper] [supplementary information]Carreira-Perpiñán, M. Á. and Renals, S. (1998):

"Dimensionality reduction of electropalatographic data using latent variable models".

*Speech Communication***26**(4):259-282.

[external link] [paper preprint] [supplementary information]Carreira-Perpiñán, M. Á. and Renals, S. (1998):

"Experimental evaluation of latent variable models for dimensionality reduction".

*1998 IEEE Signal Processing Society Workshop on Neural Networks for Signal Processing (NNSP98)*, pp. 165-173, Cambridge, UK.

[external link] [paper preprint] [© IEEE] [slides]Carreira-Perpiñán, M. Á. (1997):

"Density networks for dimension reduction of continuous data: Analytical solutions".

Technical report CS-97-09, Dept. of Computer Science, University of Sheffield, UK.

[external link] [paper]Carreira-Perpiñán, M. Á. (1996):

"A review of dimension reduction techniques".

Technical report CS-96-09, Dept. of Computer Science, University of Sheffield, UK.

[external link] [paper]

Note: this document is out of date. A much more extensive review is provided in chapters 2 and 4 of my PhD thesis.Carreira-Perpiñán, M. Á. (1995):

"Compression neural networks and feature extraction: Application to human recognition from ear images" (in Spanish).

MSc thesis, Technical University of Madrid, Spain.

[abstract] [paper] [slides] [database of ear images]

This is probably the first work that introduces the idea of ear biometrics for automatic personal identification.Carreira-Perpiñán, M. Á. (1994):

"The modelling of the thermal subsystem in spacecraft real-time simulators".

*3rd Workshop on Simulators for European Space Programmes*, pp. 69-78, ESA/ESTEC, Noordwijk, The Netherlands.

[paper] [slides]Carreira-Perpiñán, M. Á. (1994):

*ESA Thermal Analysis Program (ESATAN): An evaluation*.

European Space Agency Internal Report DOPS-SST-TN-0215-SIM, Darmstadt, Germany.

[paper]

Carreira-Perpiñán, M. Á. (2016): CSE176 Introduction to Machine Learning: Lecture notes. University of California, Merced, 2015-2016.

Carreira-Perpiñán, M. Á. (2016): EECS260 Optimization: Lecture notes. University of California, Merced, 2008-2016.

Carreira-Perpiñán, M. Á. (2006): EE 589/689 Foundations of computer vision: Lecture notes. OGI/OHSU, Fall quarter 2006.

Vladymyrov, Max (2014):

*Large-scale methods for nonlinear manifold learning*. PhD thesis, University of California, Merced.

[external link] [paper] [slides]Wang, Weiran (2013):

*Mean-shift algorithms for manifold denoising, matrix completion and clustering*. PhD thesis, University of California, Merced.

[external link] [paper] [slides]Kamthe, Ankur (2012):

*Data-driven modeling of phenomena in wireless sensor networks*. PhD thesis, University of California, Merced.

[external link] [paper] [slides]Qin, Chao (2011):

*Data-driven approaches to articulatory speech processing*. PhD thesis, University of California, Merced.

[external link] [paper] [slides]

Can't get a paper? Questions? Just email me.

Useful downloads: Adobe Acrobat Ghostscript QuickTime Windows Media Player Real Player

Miguel A. Carreira-Perpinan Last modified: Wed Feb 27 18:15:31 PST 2019