In this project we investigate whether ANN (in particular DCNN) trained on visual datasets (ImageNet) respond to reading task typical of the Statistical Learning framework express the same pattern sensitivity as found in human subject.

Dataset consists of 23 made-up characters (taken from the Brussels Artificial Character set - BACS2-serif). Each 3-character word is composed by randomly sampling the characters such that two classes emerge on character-pair distribution: high-frequency and low-frequency.

Here is an example of a randomly sample 3-character word from the BACS2-serif character set.