Stop tinkering, start measuring!
Predictable experimental design of Neural Network experiments

Epoch

Data

Capacity Demand



Bias Indicator

Which dataset do you want to use?

Features

Which properties do you want to feed in?

Double-click to remove or add link.
Click anywhere to edit.
Weight/Bias is 0.2.
Hold shift and click on two nodes to create a link between them.
The outputs are mixed with varying weights, shown by the thickness of the lines.

Memory Equivalent Capacity: 
 bits 

Output

Test loss
%
Train loss
%
Generalization:
bit/bit

Colors shows data, neuron and weight values.

Accuracy for each class
Test:
,
Train:
,

Want to give feedback?

This extension is constantly evolving, and we would very much appreciate your feedback.
The button below will open up a simple Google Form. Thanks!

Bring your own dataset

It is now possible to use your own dataset on the platform.
The dataset must be a .csv-file, with three columns x, y and label in {-1, 1}, without any headers.
An example dataset of a star can be downloaded here.

Need a tutorial video?

Click here for a Youtube tutorial video.

Further reading

Paper: From Tinkering to Engineering - Measurements in Tensorflow Playground
Cheat Sheet: Machine Learning Experimental Design Cheat Sheet
Lectures: Experimental Design for Machine Learning on Multimedia Data at UC Berkeley

Credits

This was created by Henrik Høiness, Axel Harstad, Alfredo Metere, and Gerald Friedland. For the theoretical foundations see this article. This is a continuation of Tensorflow Playground which is a continuation of many people’s previous work — most notably Daniel Smilkov, Shan Carter and Andrej Karpathy’s convnet.js demo and Chris Olah’s articles about neural networks.