Bayesian Neural Networks #141
-
|
I am new to machine learning and i would like to learn by doing. I am not a fan of python and gomlx seems perfect for me! Can this library help me implement a Bayesian Neural Network? If not, what parts need to be added to this framework? Probabilistic Layers? Perhaps I can help adding some of the missing parts, if i understand what's going on. |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 1 reply
-
|
hi Nikolas, I'm happy to hear of your interest in Bayesian NN, it's an interesting area of research. And I also agree that doing is the best way of learning. I'm pretty sure you can relatively easily build a Bayesian Neural Networks with GoMLX -- you can build any arbitrary computation graph, and there is auto-differentiation, plus supporting library to handle variable (learned weights). If you want to expose it as a layer library, consider checking out the More importantly, to get started on GoMLX, check out the tutorial (it builds the abstractions of GoMLX from the bottom up). Then browse around the examples. Feel free to ask any questions -- either here, or find me in LinkedIn. Just FYI, I'm currently working on a large PR with a new backend for GoMLX in pure Go -- more portable than the XLA backend and it will require no extra installation. But ... orders of magnitude slower. One is able to choose backends in compile and/or runtime. |
Beta Was this translation helpful? Give feedback.
-
|
Thank you for the reply! I'll take the library approach and see how far I can go.
Very interesting. Just out of curiosity, are you leveraging existing existing libraries or writing it from scratch? And regarding the performance, it should exist before it gets optimized. |
Beta Was this translation helpful? Give feedback.
-
|
Sadly, I'm writing from scratch. There are no libraries that would support the many dozens of operations a backend has to define (the Later I would love to leverage a fast matrix multiplication from external libraries, if I find some that fits -- since MatMul is the core thing that takes most time usually. One obvious candidate is gonum, it supposedly does fast matrix multiplications for float64. I'll have to investigate later. On this topic there is also the ongoing proposals to support SIMD instructions (intrinsics) in Go (latest proposal that I'm aware). I would love to leverage that for all operations once it's out. |
Beta Was this translation helpful? Give feedback.
hi Nikolas, I'm happy to hear of your interest in Bayesian NN, it's an interesting area of research. And I also agree that doing is the best way of learning.
I'm pretty sure you can relatively easily build a Bayesian Neural Networks with GoMLX -- you can build any arbitrary computation graph, and there is auto-differentiation, plus supporting library to handle variable (learned weights). If you want to expose it as a layer library, consider checking out the
ml/layers/fnnandml/layers/kanfor inspiration.More importantly, to get started on GoMLX, check out the tutorial (it builds the abstractions of GoMLX from the bottom up). Then browse around the examples. Feel free to ask any question…