Telos is a program that trains a deep learning model to learn long distance dependencies in a synthetic data set.
The data consists of two equal-length sequences. The feature sequence is a random ordering of ten digits from 0 to 9, in which one of the digits is repeated in two random places and the rest of the digits are unique. The label sequence is a list of ten 0s and 1s, where 1s appear in positions corresponding to the repeated digits and 0s appear in positions corresponding to the unique ones.
8 9 5 2 4 7 3 0 3 1
0 0 0 0 0 0 1 0 1 0
6 6 5 4 8 3 9 2 1 7
1 1 0 0 0 0 0 0 0 0
8 1 7 2 0 5 6 4 8 3
1 0 0 0 0 0 0 0 1 0
Telos trains a recursive neural network to predict the 0 and 1 labels given a sequence of digits. This requires a model to learn a simple long-distance dependency.
See telos --help
for details on generating data sets, training and
evaluating models.