Note: this page is out of date, refer to my list of papers for current work.
These are some projects I am working or have worked on:
- The method of auxiliary coordinates (MAC): a mathematical strategy to optimise "nested" systems, such as deep neural nets, without using chain-rule gradients (backpropagation), that reuses single-layer algorithms, handles non-differentiable layers and introduces significant parallelism.
- Articulatory inversion (the recovery of the vocal tract shape that produces a given speech signal), and other inversion problems such as inverse kinematics of a robot arm, recovery of facial features given speech acoustics, etc.
- Mean-shift algorithms for clustering and image segmentation (Java applet).
- Articulated pose tracking: recovering the 3D pose of a person from monocular video, using the Laplacian Eigenmaps Latent Variable Model.
- Non-rigid image registration with generalised elastic nets and other probabilistic models.
- Spectral clustering and manifold learning, with applications to computer vision, biomedical data, etc.
- Markov-chain Monte Carlo methods for estimating Markov random fields, in particular contrastive divergence learning.
- Image labelling with multiscale conditional random fields.
- Mode-finding in Gaussian mixtures: how many modes can a Gaussian mixture have?
- Continuous latent variable models for dimensionality reduction and missing data reconstruction, with applications to speech recognition and inverse problems: see my PhD thesis page.
- Generalised elastic nets: mathematical analysis, simulation and application of generalized elastic nets to cortical maps in primary visual cortex.
- Dimensionality reduction of EPG data: electropalatography is a technique to record linguopalatal contact during continuous speech.
- Ear biometrics: as part of my MSc thesis (1995), I created a database of ear images.
Miguel A. Carreira-Perpinan
Last modified: Sat Jul 9 23:45:46 PDT 2016
UC Merced |
MACP's Home Page