Bayesian neural networks

Рет қаралды 11,781

Thinkstr

Күн бұрын

My first classes at OIST are coming up! OoO
patreon.com/thinkstr

Пікірлер: 32

@anasparkle8948 2 жыл бұрын

Great video! Thank you for explaining so clearly. Do you have references for this video?

@badstylecherry7255 6 ай бұрын

Hi Thinkstr, I just came across your channel while looking up the Free Energy Principle. It's somewhat serendipitous that I managed to find your channel as I am considering studying a masters (or possibly phd) in Japan since I'm settled here. How has your experience been? Also, hope you're enjoying Okinawa, I love that place.

@Thinkstr 6 ай бұрын

Hi, thanks for watching! Okinawa is gorgeous, but I wish it had public transport like Tokyo, haha. Getting a PhD is pretty stressful, but it's usually the kind of stressful I like, working on interesting projects. I want to make another video about the FEP, but have to publish a paper first. Someday!

@HeduAI 2 жыл бұрын

That was pretty cool! Thanks for sharing :)

@Thinkstr 2 жыл бұрын

Thanks for watching! I should start making these again.

@HeduAI 2 жыл бұрын

@@Thinkstr You absolutely should :)

@ritvikmath 2 жыл бұрын

Super cool!

@Thinkstr 2 жыл бұрын

Thanks! I was told about Bayesian neural networks by someone I've met through OIST!

@bijan1316 Жыл бұрын

Great work

@Thinkstr Жыл бұрын

Thanks! These videos are a fun way for me to learn a lot.

@ZachariahRosenberg 2 жыл бұрын

Well done. Question: what was the intuition behind sticking the bayesian layer in the CNN section? The net you propose is effectively a CNN -> RNN -> Dense Classifier. The CNN primary role here is depicting the sensory input. Potentially the RNN too is just pooling sensory info, but with the inclusion of time series. The real decision making where exploration/exploitation/policy is occurring, I'd think, is toward the final most layers. What happens if you placed the bayesian layer somewhere there?

@Thinkstr 2 жыл бұрын

That's a great question! I tried doing it a few different ways, with Bayesian CNN, or Bayenian RNN, or Bayesian linear layer, or even all three of them. If I recall correctly, I presented the version with a Bayesian CNN because that was the one which ran most quickly; Bayesian RNN seemed to take the most time.

@yassinet.benchekroun5087 2 жыл бұрын

Great video!

@Thinkstr 2 жыл бұрын

Thanks! These are fun to make, but I don't have a lot of time for them nowadays.

@MLDawn 2 жыл бұрын

but why is it called Bayesian?! I thought some Bayesian update will happen?!

@williamshaw1184 Жыл бұрын

The Bayesian update happens continuously during training when the likelihood and prior are multiplied together to form the posterior of the weight distribution.

@wyldeman0O7 2 жыл бұрын

Mr Thinkstr, Question: Premise firstly - Since Convolution Neural Networks Exist, do Cross-Correlation and Auto-Correlation networks exist. If so, will you steer my imagination the correct direction. Also, still processing your vid. Cool dude you are, I know that for sure! 2ndly: has anyone build a "linguistic-interpreter-compiler" application, we need one to buffer ourselves from A.I. influence. Thanks! :D

@samuelgoldstein3590 Жыл бұрын

10/10

@Thinkstr Жыл бұрын

Thanks! These are super fun to make.

@fadouamassaoudy7832 2 жыл бұрын

make another video when things go well.........or wrong doesn't matter we liked it :)))))))))))))))))))

@Thinkstr 2 жыл бұрын

Thanks! I've been pretty busy lately, but I'll try to make at least one video this month

@ddogg9255 2 жыл бұрын

Nice demonstration. You lost me at convoluting the 3d space into 2d (I guess that's a type of transform? I have no technical bg). Always thought it was pronounced "bay-sea-anne" 😬

@Thinkstr 2 жыл бұрын

Thanks! I think I was talking too fast for that, and should've found a way to visuslise it.

@wyldeman0O7 2 жыл бұрын

Linear Algebra is Free now, learn it, you will be happy you achieved learning such a glorious thing. You will probably answer your own question too! :D