bhtucker / perceptron_viz Goto Github PK

A perceptron is the simplest form of 'neural network' learning.

The 'neuron' is simply a vector that can be multiplied against data points to say 1 or 0.

This perceptron tries to learn which letters are vowels, ie, we draw 26 points in space and say some are one class and the rest are another.

You can:

• select what letter (or groups of letters) to feed to the perceptron next
• watch it's learning pattern (initially, moving left after consonants, right after vowels)
• see the weight vector at bottom of page
• see the current accuracy as compared to a benchmark accuracy from a linear algebra class separation solution (LDA)
• adjust the dimensionality

Requires numpy/scipy (if you need these look up anaconda distribution)

Install python dependencies with

pip install .

from base directory.

Build JS app from viz directory

npm install
npm run build

After building, /voweler/dist/ should exist

Run python wsgi.py to run app simply.

Cross-request state persistence is available via shared memory managed by uwsgi: see the multiapp branch