Eric J. Ma - An Attempt At Demystifying Bayesian Deep Learning

PyData New York City 2017
Slides: ericmjl.github.io/bayesian-de...
In this talk, I aim to do two things: demystify deep learning as essentially matrix multiplications with weights learned by gradient descent, and demystify Bayesian deep learning as placing priors on weights. I will then provide PyMC3 and Theano code to illustrate how to construct Bayesian deep nets and visualize uncertainty in their results. 00:00 Welcome!
00:10 Help us add time stamps or captions to this video! See the description for details.
Want to help add timestamps to our KZbin videos to help with discoverability? Find out more here: github.com/numfocus/KZbinVi...

[@user-nk8ry3xs5u]

Great video to develop a simple mind model of neural networks. Bonus : frequentist vs. Bayesian made simple! Great work Eric!

[@harshraj22_]

1:00 Intro to Linear, Logistic regression, Neural Nets 9:40 Going Bayesian 14:32 Implementation Using PyMC3 24:27 QnA

[@mherkhachatryan666]

Love the charisma, enthusiasm put in this talk well done!

[@suzystar3]

Thank you so much! This has helped me so much with my project and really helped to understand both deep learning and bayesian deep learning much better. I really appreciate it!

[@cnaccio]

Huge win for my personal understanding on this topic. I wish every talk was given in this format. Thanks!

[@BigDudeSuperstar]

Incredible talk, well done!

[@sdsa007]

great energy! and nice philosophical wrap-up!

[@HeduAI]

Excellent talk! Thank you!

[@cherubin7th]

Great explanation!

[@bracodescammer]

I understand the benefit of modelling aleatoric uncertainty, e.g. to be able to deal with heteroscedastic noise. However, why do we need to model epistemic uncertainty? The best prediction after all, lies in the middle of the final distribution. If you sample from the distribution, you will lose accuracy. So is uncertainty only useful for certain applications to determine different behaviour based on high uncertainty? For example: If uncertainty is high, drive slower?

[@catchenal]

The other presentation Eric mentions is that of Nicole Carlson: Turning PyMC3 into scikit learn kzbin.info/www/bejne/sHi1n5yol62KgJo

[@vtrandal]

Point #1 is wrong. You left out activations.

[@bonob0123]

The tanh and Relu nonlinearities are the activations. He is not wrong. You are wrong. Learn to be humble.

[@MiKenning]

Was he referring to Tensorflow when he denigrated an unnamed company for its non-pythonic API? The new Tensorflow is much better!