GSoC_2015_project_fundamental

Fundamental Machine Learning Algorithms, Part 2

Linear Latent Gaussian Models: Linear and non-linear

Mentors

• Fernando (github: iglesias, IRC: iglesias)
• Heiko (github: karlnapf, IRC: HeikoS)

Difficulty & Requirements

Easy/medium/advanced. You need to be able to

• get confused by C/C++
• trim beards
• count socks

Description

Models: State space dynamics are linear Gaussian Observation links might be Gaussian but can also be non-linear. Both learning and inference.

Models:

• LGSSM
  • Kalman filter for inference
  • EM for learning
  • HO-Kalman for fast EM learning
• Non-linear GSSM: Let's focus on exp-family link functions
  • Inference: variational, EP for the E-step
  • Learning: variational EM, See Büsing & Macke "Non-linear Kalman"

Other ideas:

• Use existing EM framework for the variational apprixmate methods
• Could be that we need to change it, maybe even another iteration on GMMs
• HMM re-implementation. Viterbi, Baum Welch, all neat and clean? Notebook?
• Other Linear Gaussian models: FA, PPCA, also work with the above algorihtms (EM, gradient descent)
• ARD for LGMs. Prior on columns of projection matrix, compute posterior

Details

Write about details of the project here.

Waypoints and initial work

• More Random forests (BART)
• Please put your ideas here

Optional

Parts of the project that would be cool once the core is finished.

Why this is cool

Motivation to get involved here.

Useful ressources

• Put a list of ressources/links here