變分推斷 貝葉斯神經網路有什麼論文可以推薦閱讀嗎?
我現在處於本科階段,對這方面挺感興趣的,想以後進行相關方面的研究。
希望大家能推薦一下這方面的論文。比如變分推斷的整個發展過程從meanfild 到 vae 再到flow auxiliary 之類的。。如果能加上推薦閱讀順序就更好啦。
貝葉斯神經網路相關方面的也可以
現在比較焦慮。感覺完全沒有方向,並且功利心好像有點嚴重,急切想發論文。不過還是要腳踏實地地一步步來,希望大家能推薦一下相關領域的論文。
謝謝大家
除了樓上答案推薦的David Blei的Variational Inference: A Review for Statisticians 之外,感覺當時課上module 4的reading主要是required一部分很值得一看,只能說Eric Xing是個狠人...(對於我這個數學渣而言_(:з」∠)_...)
10708 Probabilistic Graphical Models?www.cs.cmu.edu搬運如下...
Lecture 12 (Eric) -Slides
- Variational Inference: Loopy Belief Propagation
- Ising models
Required:
- Yedidia et al., Generalized Belief Propagation
- Wainwright and Jordan, Graphical Models, Exponential Families, and Variational Inference, Sec. 4.1
Optional:
- Murphy et al., Loopy Belief Propagation for Approximate Inference: An Empirical Study
Lecture 14 (Eric) -Slides-Video
- Theory of Variational Inference: Inner and Outer Approximation
Required:
- M. Wainwright and M. Jordan, Variational Inference in Graphical Models: The View from the Marginal Polytope
Optional:
- M. Wainwright and M. Jordan, Graphical Models, Exponential Families, and Variational Inference, Sec. 3 and Sec. 4
Lecture 13 (Willie) -Slides-Video
- Variational Inference: Mean Field Approximation
- Topic Models
Required:
- Xing et al., A Generalized Mean Field Algorithm for Variational Inference in Exponential Families
- Wainwright and Jordan, Graphical Models, Exponential Families, and Variational Inference, Ch. 5 (Section 5.1 - 5.3)
Optional:
- Variational inference tutorial (NIPS 2016) (同樣來自David Blei,其實這個tutorial相當不錯)
個人覺得,VI最重要的就是推ELBO,然後Amortized VI加一些trick去SGD。在graphical model里是一種工具性質的存在。
文章實在是很多,但是近幾年的工作基本上就是上面這一個套路。推ELBO你可以去看「Variational Inference A review for statisticians」,不管mean field還是AVI都是一個推法。AVI你隨便找一篇最近的文章看一下就知道了,無非就是reparameterization trick或者score gradient.
另外,你說的這兩個東西我老闆學生時代(20年前)就在做了,純做這個在現在應該不屬於容易發論文的範疇。
一些貝葉斯神經網路的paper
Expectation propagation:
- Probablistic backpropagation
Variational inference:
- Bayes by backprop
- Local reparameterization trick
- Dropout as bayesian approximation
- Concrete dropout
MCMC:
- SGLD
- preconditioned-SGLD
- SGHMC
Particle approximation:
- Stein variational gradient descent
robi56/awesome-bayesian-deep-learning?github.com
我建議先看Bayes-by-backprop和SGLD,以及Dropout as bayesian approximation,因為最簡單。SGLD就是SGD+高斯雜訊,MC-dropout就是test-time繼續dropout……
不過,做BNN的話,感覺mean-field的posterior不太好用,ELBO里的KL-divergence項非常容易over-regularize