Simply wonderful! The content is up to date and covers every aspect of Deep Learning I can think of. (I myself have been doing data mining teaching/research for the past decade, but my knowledge in deep learning needs constant updates.) I can't wait to learn from the yet-to- release lectures. Thank you!
I have worked and read many resources on Bayesian Deep Learning, but your presentation is far beyond anything that I have seen. Super clean, intuitive, and easy-to-follow. So Thank you :)
Excellent coverage of the topic. However, a discussion of how easy or difficult it is to engage in evidential deep learning for real problems, fast algorithms as well as a discussion on computational complexity would be immensely useful.
such a difficult content explain in easy to understand manner. To make highly complex thing very easy to understand you really need to have very in-depth understanding of subject and also need to understand students mentality ( thinking from students perceptive what they can understand and how to make explanation as simple as possible) . Both lecturer of this course really great in both of this two things
Thank you for this power packed lecture. Literally every sentence was a knowledge for me! Where can find the slides of this lecture? I couldn't find slides on the link provided in description. If you provide me link to download the slides, that would be awesome. Thank you so much again!!
anyone has any idea how to reproduce Fig. S3 from the paper? I don't understand this part: "Rather than using the L1 error in the regularization term, as in previous experiments, we use regularize the standard score and estimate epistemic and aleatoric uncertainty (Fig. S3). " I think in the previous experiment, epistemic always higher than aleatoric uncertainty (see Fig 3), but magically it got sorted as shown in Fig S3.
Is there an introduction or a book and some github python code about evidential deep learning someone can share? Is there a difference between bayesian evidential learning and eveidential (deep) learning?
Definitely!! Evidential layers can be placed at the end of an LSTM (for example) to model the uncertainty at each timestep (for many-to-many problems) or at the final timestep (for many-to-one problems).
Exactly, for any supervised problem (e.g., trained with MSE loss for regression or cross entropy for classification) it should be a simple drop-in replacement to use evidential layer/loss instead. If you don't have a supervised learning problem, you can also obtain uncertainty estimates using sampling based techniques and treating your model as a Bayesian NN (e.g., by using Monte Carlo Dropout sampling).
Hello, I don’t get exactly how you calculate the higher moment. Do you output E(X^n) on top of E(X) from a deterministic RNN ? Thanks a lot for the really good video
Hi, thanks for this really interesting content ! I was wondering how uncertainty (aleatoric and epistemic) can be related with the Bias/Variance Decomposition : _Expected Test Error = Variance + Bias² + Error_ I would say that : (i) *aleatoric uncertainty = Error* (i.e the inherent noise in the training dataset, due to noisy data or wrong labels) (ii) *epistemic uncertainty = Variance* (i.e "how far my prediction - obtained with one particular training dataset - is from the average model obtained in theory with all possible data") (iii) *Bias is independent from uncertainty* Not sure about (iii) ... Actually, not sure about everything ! Would be nice to have your opinion on that ! =) Thanks a lot.
The more I think about this, the more I think this is not a good idea. You're asking the computer who other than cats and dogs hasn't seen anything to tell you how sure it is what it is looking at. But there are many cases where something will look more like a cat or more like a dog, for example a city with two houses that look like cat ears. My point is it may say it's very sure it's more like a cat but it can't tell us if it's a cat because it didn't learn what a cat is, it only learnt how a cat is different from a dog.
Simply wonderful! The content is up to date and covers every aspect of Deep Learning I can think of. (I myself have been doing data mining teaching/research for the past decade, but my knowledge in deep learning needs constant updates.) I can't wait to learn from the yet-to- release lectures. Thank you!
I have worked and read many resources on Bayesian Deep Learning, but your presentation is far beyond anything that I have seen. Super clean, intuitive, and easy-to-follow.
So Thank you :)
Just getting to know about this channel..
I'm a complete beginner and I hope I can make use of this
Excellent coverage of the topic. However, a discussion of how easy or difficult it is to engage in evidential deep learning for real problems, fast algorithms as well as a discussion on computational complexity would be immensely useful.
This one is my personal favorite giving that I want to understand your latest ICRA paper, thank you.
Thanks!
Amazing lecture, I cant wait for 2022 videos.
Very high-quality presentation!
such a difficult content explain in easy to understand manner. To make highly complex thing very easy to understand you really need to have very in-depth understanding of subject and also need to understand students mentality ( thinking from students perceptive what they can understand and how to make explanation as simple as possible) . Both lecturer of this course really great in both of this two things
Thank you for this power packed lecture. Literally every sentence was a knowledge for me!
Where can find the slides of this lecture? I couldn't find slides on the link provided in description. If you provide me link to download the slides, that would be awesome. Thank you so much again!!
Concise and clear, great lecture. Thanks for making it available to general public.
There may be a typo in the slides, the paper by Sensoy was published in NeurIPS2018, not NeurIPS2019.
Thank you.
A great, clear lecture!
very informative.thanks to Amini sir
anyone has any idea how to reproduce Fig. S3 from the paper? I don't understand this part: "Rather than using the L1 error in the regularization term, as in previous experiments, we use regularize the standard score and estimate epistemic and aleatoric uncertainty (Fig. S3). " I think in the previous experiment, epistemic always higher than aleatoric uncertainty (see Fig 3), but magically it got sorted as shown in Fig S3.
Is there an introduction or a book and some github python code about evidential deep learning someone can share?
Is there a difference between bayesian evidential learning and eveidential (deep) learning?
This might be helpful: github.com/aamini/evidential-deep-learning
Great lecture, very clear.
can already see how this is going to positively effect medical imaging :))
Evident Training se Doctor Bimari or uski Medicine sahi se Predict kr payege agar ye Deep Learning Tech hospital mai Use kre 👈🏼🤌🏼👌🏼
hi. thanks for this great lecture.
can I get the implementation code of this algorithm?
Yes! Here you go: github.com/aamini/evidential-deep-learning/
There is a mistake at 9:15. It should allow for 'bases different from e' by putting (e^(beta*z)).
i am working on a variation of deep evidence regression, and love the intuition in this lecture !
Wonderful lecture! Thanks 👍
Great lecture!
Exceptional!!
Great lecture!Thank you very much!
Loved every bit of the lecture! :)
Just great!
Is this possible for time series analysis?
Definitely!! Evidential layers can be placed at the end of an LSTM (for example) to model the uncertainty at each timestep (for many-to-many problems) or at the final timestep (for many-to-one problems).
@@AAmini so basically I transform my tiempo series to a supervised learning problem (x,y) to run this algorithm?
Exactly, for any supervised problem (e.g., trained with MSE loss for regression or cross entropy for classification) it should be a simple drop-in replacement to use evidential layer/loss instead. If you don't have a supervised learning problem, you can also obtain uncertainty estimates using sampling based techniques and treating your model as a Bayesian NN (e.g., by using Monte Carlo Dropout sampling).
@@AAmini And the return of this algorithm is the mean, variance as well as the prediction of my time series, right?
Hello, I don’t get exactly how you calculate the higher moment. Do you output E(X^n) on top of E(X) from a deterministic RNN ? Thanks a lot for the really good video
It's excellent lecture. Thank you so much.
Woow..Awesome Lecture!
Thanks, great lecture!
Hi what is the constraints of alpha,beta,lamda and gamma?
Is f(X) of epistemic variance f(x|w_t) ? Also, f(x|w) means y?
Hi, thanks for this really interesting content !
I was wondering how uncertainty (aleatoric and epistemic) can be related with the Bias/Variance Decomposition :
_Expected Test Error = Variance + Bias² + Error_
I would say that :
(i) *aleatoric uncertainty = Error* (i.e the inherent noise in the training dataset, due to noisy data or wrong labels)
(ii) *epistemic uncertainty = Variance* (i.e "how far my prediction - obtained with one particular training dataset - is from the average model obtained in theory with all possible data")
(iii) *Bias is independent from uncertainty*
Not sure about (iii) ... Actually, not sure about everything !
Would be nice to have your opinion on that ! =) Thanks a lot.
No way there’s anyone who was able to understand the “evidential learning” part without looking into other videos or having prior knowledge.
The more I think about this, the more I think this is not a good idea. You're asking the computer who other than cats and dogs hasn't seen anything to tell you how sure it is what it is looking at. But there are many cases where something will look more like a cat or more like a dog, for example a city with two houses that look like cat ears. My point is it may say it's very sure it's more like a cat but it can't tell us if it's a cat because it didn't learn what a cat is, it only learnt how a cat is different from a dog.
Evidence Training is there to overcome that Problem
So that System can take Evidence from Outside world and Learn it from themselves
Hello, I have a question, what is ~ in this formula for example y ~ Normal(miu, sigma^2)?
Thank you for the nice lecture! :)
~ 👈🏼this sign means its almost Equal to
= 👈🏼 this is Equal to Sign
~ 👈🏼this is ALMOST(about) Equal to sign
Great....
可以在标题上标明第几集吗?
This lecture slides are not avilable on the website.
Please use the archived link. All content from past years is still archived online. Thanks!
@@AAmini Thanks!
Sir, I have watched first 2 lecture. I am struggling to generate code and models. Kindly help me with practicals 😌
Lecture on Recurrent Networks and CNNs ?
Bhai bina Coding k IT job mil sakti hai kya???
Mujhe Coding bilkul Nahi pasand
is this lecture for bachelor degree? or magister?
it is master *