shanest/color-learning

Accompanies the paper, "Ease of learning explains semantic universals". Generates and trains neural networks on artificial color systems, showing that _degree of convexity_ explains accuracy.

Ease of Learning Explains Semantic Universals

This repository accompanies the paper:

• Shane Steinert-Threlkeld and Jakub Szymanik, "Ease of Learning Explains Semantic Universals", under revision at Cognition. https://semanticsarchive.net/Archive/zM5ZGIxM/EaseLearning.pdf

We generate artificial color naming systems by partitioning the CIELab color space, then train neural networks to learn those systems. We show that the ability of networks to learn these color systems is well-explained by their degree of convexity, supporting a semantic universal for color terms across language.

Requirements

Python 2.7, TensorFlow 1.10, Numpy 1.13, Pandas

(NB: the code should be compatible with Python 3, but has not been tested with it. Later versions of TF and NP will also likely work, but no promises.)

Running Experiments

Calling

python run_experiment.py

will run the main experiment reported in the paper. In the output directory trial, the following will be output:

• results.csv: each row is one trial, recoring the color system generating algorithm's parameter values, degree of convexity, network accuracy, as well as other geometric variables, and the actual partition
• points.npy: a numpy array, with the CIELab space points

To see how an experiment is defined and to play with defining your own, you can check out the main_experiment method, which itself repeatedly calls run_trial.

The algorithm for generating artificial color naming systems (and some utility methods for it) are in partition.py.

Analyzing the Data

The regression and commonality analysis reported can be run via

python analysis.py

This will also produce trial/complex_regression.png.

The cluster analysis is handled by clusters.R, an will produce the cluster PNG files.