https://github.com/bddean/BQNprop

Toy backpropagation implementation written in BQN.
https://github.com/bddean/BQNprop

Last synced: about 2 months ago
JSON representation

Toy backpropagation implementation written in BQN.

• Host: GitHub
• URL: https://github.com/bddean/BQNprop
• Owner: bddean
• Created: 2021-06-26T23:44:20.000Z (about 3 years ago)
• Default Branch: main
• Last Pushed: 2021-06-27T03:11:58.000Z (about 3 years ago)
• Last Synced: 2024-05-08T09:40:12.782Z (5 months ago)
• Size: 2.93 KB
• Stars: 3
• Watchers: 3
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-bqn - BQNprop

README

        
# BQNprop

BQNprop is a toy backpropagation implementation I wrote to learn about

machine learning and array programming. It's extremely bare-bones:

Hyperparameters are all hardcoded, there are no bias nodes, and I haven't

even tested this on any real data.

I am still pleasantly surprised at how short the implementation ended

up being. Only 17 lines excluding comments, blank lines and testing.

# Future work

Real ML libraries apparently use Automatic Differentiation (AD)

nowadays. I've been looking into http://conal.net/papers/essence-of-ad/

as a potential implementation approach, which involves compiling functions

to an efficiently differentiable form. For BQN this might mean a codegen step.

And ideally we'd want BQN running on a GPU for this.

# References

I used [3Blue1Brown](https://www.youtube.com/watch?v=Ilg3gGewQ5U&t=750s)

and [Michael Nielson's online

textbook](http://neuralnetworksanddeeplearning.com/chap2.html)

to learn the math. I used [this worked-out example

](https://steemit.com/ai/@ralampay/training-a-neural-network-a-numerical-example-part-1)

as a test.

And, of course, the [BQN docs](https://mlochbaum.github.io/BQN/).