Thank you :)

Thank you!

Thanks Arjun!

Cheers!

That's great!

Thank you! Appreciate the feedback :)

Thank you!

Thank you!

Thanks for the suggestion - the puffing sound has been a real pain! We'll check out the pop filter.

Thanks, glad you liked it!

Thank you :)

Thank you!

If I had to choose one name, I would call it machine learning (which is a subset of artificial intelligence): mathematics are the fundamental tools, statistics offers analytical frameworks and computer science is a whole load of fun things :). But practically speaking, you will come across neural networks at college if you take a machine/deep learning course as part of a computer science / information engineering degree. However, you can find lots of excellent online resources to introduce you to the world of machine learning before you go to college.

Thanks!

Thanks! Yes you're correct - averaging an ensemble of models can be viewed as assigning equal importance to each model seed rather than weighing by the posterior distribution of weights.

@@TwinEdProductions there are quite a few Bayesian package out there, like pymc3, gpytorch... it would be great if you have a tutorial to compare the features and show some hands on applications for these packages. And show the benefit of Bayesian. I’m assuming the gpytorch model will require more computing power. Thanks again.

@Canal fishing Thank you for the suggestion - we will add this to our ToDo list of videos! I'm personally quite a fan of BLiTZ as a Bayesian Neural Network library for PyTorch while I believe gpytorch is a library for Gaussian Processes.

@@TwinEdProductions you are right, I will check it out, thank you.

Thanks!

Thanks! Agreed, KL is not strictly a distance measure but helps intuition to think of it as one :)

Hi Walter, I must say you have asked the question which had bugged me for a couple of years as there seemed to be a mismatch between practical deep learning and the theoretical optimisation problem. And for some reason a lot of deep learning practitioners are not even aware of the link between the theoretical optimisation problem and the practical loss functions such as cross entropy loss. So, I have typed up a proof for both regression and classification to explain from first principles why we use cross-entropy loss. Here is the link to it: drive.google.com/file/d/1EJF1vlwD3QdyEAcrFc26JMxs4LgRYOnR/view?usp=sharing

Oh and also, the question is not banal at all!

Thanks for the support! Hard to pinpoint a single reference but probably mostly the material I had learnt during the 'advanced machine learning' course at my university.

Hi, if you are familiar with building typical neural networks in pytorch, then the BLiTZ library is a very straigtforward way to build a BNN for any equivalent network that you are interested in such as regression. The link of a tutorial introducing the library is here: towardsdatascience.com/blitz-a-bayesian-neural-network-library-for-pytorch-82f9998916c7

Hi Jahid, would you be able to reference the context in which you have come across the terms 'mean aggregator' and 'burning' as they can have different interpretations within different applications. For example, for 'burning' are you referring to 'burn-in' time in the development of statistical systems?

Yes! In fact that is a very common application of bayesian neural networks

@@TwinEdProductions but in my training the RMSE values is really high after I am testing the the trained network using bnn. I am predicting it for 500 samples to get mean but the error is high

@@user-bl2ht7bu4u does training work effectively when using a standard neural network for regression? And what is your motivation for using a BNN in your use case?

Thanks for the idea! Yes, a python example from scratch would be nice

Aha, possibly!

Hi! The video here is simply explaining the theoretical concept of Bayesian neural networks. If you're interested in the implementation side, simply check out great packages like BLiTZ (A Bayesian neural network library for pytorch).