A project to compare human intuition agains neural networks with TensorFlow
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
server
ui
.gitignore
README.md
hidden_layers_mixed.py
hidden_layers_one.py
hidden_layers_two.py
humanIntuitionUtils.py
humanIntuitionUtils.pyc

README.md

Human Intuition Challenge

Webapp to deliver human intuition challenge as well as collect data that can be used to train a neural network

You can view the current article that discuss this challenge here:

Are you Intuitive? Want to Challenge my Machine?

You can view a discussion about the creation and training of the neural network code here:

How to Design Neural Nets for the Human Intuition Challenge

Any feedback or extensions of this work is greatly appreaciated!

Start Up NodeJS Web App

You will need to have NodeJS and NPM installed (NPM comes with newer versions of NodeJS)

cd server
npm install
node app

Navigate to: http://localhost:3001/

Switch View Modes in Webapp

The web application can be in one of two modes: "view" or "edit"

"view" mode is used to deliver the human intuition test to people without them having the ability to edit any of rule sets or generate data. It is basically a locked down version of the app.

"edit" mode could be considered admin mode where you can view/modify the rule sets and generate data for machine learning

To change modes simply hit one of the following urls:

http://localhost:3000/services/mode/edit

http://localhost:3000/services/mode/view

You should see a json response of success and then you can navigate back to the homepage

Collecting Human Results

Each time someone completes the online intuition test, it will store the results summary in memory on the server. You can retrieve the json array of results by hitting this endpoint:

http://localhost:3000/services/results

Restoring the Default Schemes

Feel free to play around with the schemes, add more, modify the transformations, go nuts!

If you want to restore the schemes to the default A and B schemes used in the Human Intuition Challenge navigate to this endpoint:

http://localhost:3000/services/setdefaultschemes

Generating Data For Machine Learning

Once you have the rule sets setup to generate random strings for each scheme, you can generate as much data as you want for machine learning.

First, make sure you are in "edit" mode. This will expose a "generate data" button on the homepage (see instructions above to get the webapp into edit mode)

Simply click the "Generate Data" button and the server will spit out a file in the 'server/exports/' directory.

This file is the file the TensorFlow code is looking for when attempting to train.

Changing the amount of data generated for machine learning

If you want to increase the amount of data generated, you need to modify a parameter in the UI code.

Find the ui/scripts/modules/MachineLearning/services/DataGeneratorService.svc.js file

Then change the "DATA_POINTS_PER_SCHEME" variable in that file.

Run Machine Learning

This code was written using Python 2.7.12. You will need to have that installed as well as the version of TensorFlow that supports that version of Python.

There are several neural networks that can be trained using the data exported from the previous steps above.

Single Hidden Layer

This neural network has a single hidden layer with no activation function

python hidden_layers_one.py

Two Hidden Layers

This neural network has two hidden layers with Relu activation functions

python hidden_layers_two.py

Combined Single Hidden Layer and Two Hidden Layers (mixed)

This neural network was built to try to take advantage of all possible rules that can be used to generate strings. I found this one to work the best so far. The inputs get sent into two different channels, one a single hidden layer with no activation function and one channel that has two hidden layers with Relu. The outputs from those channels are merged via addition.

python hidden_layers_mixed.py

Start Up Tensor Board

TensorFlow comes with an awesome setup for visualizing the graph of the network and analysis of variables. All three neural networks above currently export "accuracy" that can be viewed in TensorBoard. Simply run the neural networks so the log files get created, then run this command:

tensorboard --logdir=logs