[DeepBayes2019]: Day 5, Lecture 3. Langevin dynamics for sampling and global optimization

[@miladaghajohari2308]

great presentation. It gave me a lot of insight especially given the current developments in score-based models and generation with them.

[@leonardohuang]

Fantastic talk! Clear, informative and vivid.

[@Tryndamereization]

Very good presentation. Thank you

[@jfndfiunskj5299]

Very good. Well done.

[@sayarerdi]

Could you also share the codes of your animations?

[@qiuhaowang2063]

Helpful talk! Thank you . I just don't understand why there is a 'div' in the equation at 16:45. Could you show me some references?

[@k_neklyudov]

Remember that div(f) is just a convenient notation for \sum_i df_i/dx_i. If you'll try to take the jacobian of the formula for x' you will end up evaluating the determininant of a large matrix, and there you should see that some terms of this determinant are negligible: o(dt). Thus, in some sense, the diagonal of the matrix (which is 1+df_i/dx_i*dt) plays the main role and, thus, div appears.

[@ivanmedri288]

I know it is old post but could you solve this? I am stuck on the same part.

[@abhishekmaiti8332]

At 34:01, are \thetas parameters of the model or the sampled data points from a model which is parameterised by \theta?

[@k_neklyudov]

thetas are the parameters of the model, which could be a regression or a classifier in this case. hence, here we sample the parameters of a model (say weights of linear regression), and then we should average the predictions over all these sampled parameterrs.

[@abhishekmaiti8332]

@@k_neklyudov I see, thanks a lot for the clarification.

[@nadineca3325]

Thank you for this informative discussion and amazing presentation!

[@PradeepBanerjeeKr]

Simply super!

[@hakiim.jamaluddin]

Super clear, thanks!

[@MrDudugeda2]

great talk!

[@해위잉]

Fantastic,,,, i want to go samsung

[@umountable]

Maybe its useful to motivate a little more before showing only slides packed with equations for half an hour straight

[@MNasirAziz]

I understand you. But you can look who he uses final derived equation by just skipping derivation.

[@k_neklyudov]

Hi Stefan! Thanks a lot for your feedback! I totally agree that this talk taking separately lacks motivation and practical examples. The reason is that this talk was designed as a part of Day 5, which is about MCMC. You can also find the discussion of the Langevin dynamics in this talk kzbin.info/www/bejne/h5ClmmV-brN-sMU.

[@GenerativeDiffusionModel_AI_ML]

very good talk!