analysis-by-synthesis by learning to invert generative black boxes, wip

19 April 2017

notes on (Nair et al., 2008).

understanding: 5/10
code: ?

task

given:

• training data \(\{y_n\}\)
• black box generative model \(p(y \given x)\)

goal: approximate posterior \(q_{\phi}(x \given y) \approx p(x \given y)\). (not exactly posterior since they have no defined prior).

approach

1. pick \(y_n\) from training data set.
2. run it through the recognition and perturb to obtain \(x\).
3. obtain \(y\) from black box generator.
4. perform supervised learning on \((x, y)\):
   1. get \(x'\) from recognition network.
   2. calculate loss between \(x\) and \(x'\).
   3. take gradient step.

---

references

1. Nair, V., Susskind, J., & Hinton, G. E. (2008). Analysis-bySynthesis by Learning to Invert Generative Black Boxes.
   • Bib

   @inproceedings{nair2008analysis,
     title = {Analysis-bySynthesis by Learning to Invert Generative Black Boxes},
     author = {Nair, Vinod and Susskind, Josh and Hinton, Geoffrey E.},
     year = {2008}
   }
   

[back]