4 ms·

> I can probably write vanilla JS or Python code that does that in 10 lines or so and would finish in around 5 seconds. Really? There are certainly frameworks

> I can probably write vanilla JS or Python code that does that in 10 lines or so and would finish in around 5 seconds. Really? There are certainly frameworks that you can use to achieve this performance in that amount of code, but I would be really interested to see that in vanilla JS/Python.

rawnlq 10y ago
Not op but here's a cool example in JS: http://cs.stanford.edu/people/karpathy/convnetjs/demo/mnist.html http://cs.stanford.edu/people/karpathy/convnetjs/demo/mnist....
inglor 10y ago
Hey, sure - here's 9 lines of code that implement a perceptron. I get ~95% of precision after a second with 400 samples. ```js const dotSign = (v1, v2) => v1.reduce((prev, x, i) => prev + x * v2[i], 1) > 0 ? 1 : -1 module.exports = (data, weights = Array(data[0].content.length).fill(0)) => { for(const {label, content} of data) { const delta = (label - dotSign(content, weights)) / 2; weights = weights.map((x, i) => x + delta * content[i]); } return { perceive: vector => dotSign(vector, weights), weights }; } ``` I can upload an electron app that does this with mnist if interested.
- anentropic 10y ago
  how long does does it take to get from 95% -> 98.8% though?
  - lunula 10y ago
    This. You can also get near 98% accuracy with vw and one quadratic interaction over the pixel space. It takes seconds. But that last .8%, forget about ever getting there with infinite training time.
  - inglor 10y ago
    About 10 seconds and 10 times the samples does that.
- attractivechaos 10y ago
  > I get ~95% of precision after a second with 400 samples. MNIST has 60k training samples and 10k test samples. Are you using only 400 of them? Is 95% the accuracy on the test samples or on the same set of training samples? I believe when we talk about MNIST accuracy, we always refer to the accuracy on the 10k test samples.
- quinnftw 10y ago
  Is this not a binary perceptron? How do you plan to classify all 10 digits with this?
  - inglor 10y ago
    I cheated here (this _is_ a binary perceptron), but the conversion is super easy with one-vs-rest or one-vs-one (training all pairs and then doing pairwise comparison until we find the one that matches), that would still take ~20 seconds to train.

rawnlq 10y ago

Not op but here's a cool example in JS: http://cs.stanford.edu/people/karpathy/convnetjs/demo/mnist.html http://cs.stanford.edu/people/karpathy/convnetjs/demo/mnist....

inglor 10y ago

Hey, sure - here's 9 lines of code that implement a perceptron. I get ~95% of precision after a second with 400 samples. ```js const dotSign = (v1, v2) => v1.reduce((prev, x, i) => prev + x * v2[i], 1) > 0 ? 1 : -1 module.exports = (data, weights = Array(data[0].content.length).fill(0)) => { for(const {label, content} of data) { const delta = (label - dotSign(content, weights)) / 2; weights = weights.map((x, i) => x + delta * content[i]); } return { perceive: vector => dotSign(vector, weights), weights }; } ``` I can upload an electron app that does this with mnist if interested.

anentropic 10y ago
how long does does it take to get from 95% -> 98.8% though?
- lunula 10y ago
  This. You can also get near 98% accuracy with vw and one quadratic interaction over the pixel space. It takes seconds. But that last .8%, forget about ever getting there with infinite training time.
- inglor 10y ago
  About 10 seconds and 10 times the samples does that.
attractivechaos 10y ago
> I get ~95% of precision after a second with 400 samples. MNIST has 60k training samples and 10k test samples. Are you using only 400 of them? Is 95% the accuracy on the test samples or on the same set of training samples? I believe when we talk about MNIST accuracy, we always refer to the accuracy on the 10k test samples.
quinnftw 10y ago
Is this not a binary perceptron? How do you plan to classify all 10 digits with this?
- inglor 10y ago
  I cheated here (this _is_ a binary perceptron), but the conversion is super easy with one-vs-rest or one-vs-one (training all pairs and then doing pairwise comparison until we find the one that matches), that would still take ~20 seconds to train.