OCR and Neural Nets in JavaScript | Hacker News Reader

OCR and Neural Nets in JavaScript

ejohn.org

66 pointsoption_greek18 comments

I'm writing an ML in JS series and I was excited to see another person talking about neural networks in Javascript (an upcoming post of mine). I was a little disappointed to see that the article doesn't talk about the actual neural net!

But still the message is clear: Javascript can do some damned cool things.

It's a turing complete language. Why is it surprising or note-worthy that you can implement ML algorithms in it?

Because up until relatively recently, JS was slow and crippled. The public image of JS is that its only use is Ajax and making popups. It's only recently being recognized as a "real language". Surprising and note-worthy that you can do ML in JS? No. But novel? Yes.

That you can implement them isn't impressive.

But most ML algorithms ("applying a pretrained neural network" being a notable exception to that) are compute-intensive, so in a language that doesn't have access to Lapack/Blas/etc, performance will be terrible.

This article was from 2009 too, and since then typed arrays have arrived which HUGELY sped up my messing-around-with-neural nets (without me actually having to write good code)!

link?

http://burakkanber.com/blog/

I've recently started a Physics in JS and a ML in JS series. Hope you like it!

I really don't know why people are so surprised/amazed by this kind of stuff. As soon as JavaScript had access to pure image data in the browser there was nothing else holding it back from these kinds of algorithms. It was always the sandbox that was holding it back, not the language.

Cool none the less...

Bear in mind this post is from 2009 so it was a bit more amazing back then ;-)

Related: https://www.coursera.org/course/neuralnets starting on 24 September 2012.

Does anyone here will do this class?

I'm currently doing the ML class, so I'm not sure.

Signed up. I'm doing algorithms-1 course now - so not sure if I can complete all assignments. Btw, how is the ML course ? Is it too math intensive...

The ML class is pretty math heavy, but he walks you through 95% of it

I'm on Week 3 and so far the math wasn't intensive. As someones said, he walks you through it.

I'm taking both of those also, and the math in ML isn't intensive... but there is much more exposure to calculus and linear algebra in ML than in Algorithms I, and while I agree with the other commenter who said Andrew Ng walks you through it, an understanding of both helps with getting the "why" of things like the gradient descent formulas introduced in the linear regression lectures.

The ML course does start out with an optional linear algebra review lecture, but it's very focused on mechanics rather than underlying mathematical reasoning... for conceptual mastery (or even just brushing up) I'd recommend going through Khan's videos on the subject, if you have the time.

I've been working on a basic multi-layer perceptron network using IndexedDB. I found IDB performance is not quite as fast as needed to pull it off with the algos I was using

Is there any specific reason you are using a DB in this case ? As I understand, for most of the cases, the neural net data stays in memory and is persisted to disk only to save and load the network.

For me the big win is if this can be done purely on the client side. I see two main advantages: 1) the user doesn't have to give up her privacy in exchange for recommendations 2) the service doesn't have to burn CPU it would have had to otherwise

When you train your network, you'll want to be loading mini-batches into memory. I've found flat files just fine for data sets of about 50,000 to 300,000 samples. The model itself will just be a matrix and a vector for each layer. Feel free to email me if you have any questions about getting going/speeding things up, happy to help. They can be really slow or reasonably fast depending on the implementation.

Reply on news.ycombinator.com