tlienart / thox

[done] phd thesis @ oxford stats

0 stars 0 forks source link

TASKS #20

Closed tlienart closed 6 years ago

tlienart commented 7 years ago

Part 1

Background

Add brief background on generic MCMC methods (maybe follow the tour by Green and Lutzysnzki)
- MH
- Gibbs
- HMC

PDMP Part

Expose results, clarify why useful
Do experiments as suggested by Leonard

Part 2

EP for Distr Bayes

Add experiments for Logreg and explain a bit more. Maybe take MAP estimator via SGD as baseline but STAN is also ok.
Suggest that Bayesian LogReg is in fact not a very good use case and that we'll come back to that in Conclusion

Intro and Conclusion

Conclusion

Discuss Bayesian ML, question of n versus p and cite loads of paper
Suggest that it seems to us that best way forward is to consider complex models with a sparse graphical model with few observations per node and that's where Bayesian would still be relevant (PDMP is a good way forward)