Off We Go!

The compile() method then readies the learning model for processing, determines mean_squared_error (deviations from optimal learning success are measured according to the mean square method) as the learning parameter, and sets rmsprop, a common method for neural networks, as the algorithm optimizer.

Line 45 then calls the model's fit() method, passes the learning data to it, and stipulates the number of learning iterations as epoch = 500. In my lab, a lower number affected the results negatively, but larger values of epoch did not achieve any greater learning success, since the system reached a steady state afterward, and the learning process visibly stagnated at a constant value for the loss function.

The section starting in line 48 of Listing 1 then monitors the predictions with the model in the current phase in the training, both for the training data and for previously unseen sequences that the system needs to guess without precedence in the training data.

Figure 7 shows that the network does a good job, even for cyclic data (1,2,3,1,2,3), regardless of the length of the signal period, which is a major advantage over traditional neural networks, for which you have to state the periodicity in advance for reliable predictions.

Figure 7: In a cyclic sequence, the algorithm predicts the next values more or less correctly.

Fast as a Snail

You need to install the following libraries from the Python treasure trove to be able to install the keras library:

pip3 install --user keras pandas tensorflow sklearn numpy
sudo apt-get install python-tk

The TensorFlow back end used by keras is not exactly speedy; it took a good 10 seconds for the program to get going and start the training on my five-year-old PC.

At the end of the day, the network did not perform particularly well in the intelligence test (Figure 8; see also the "Online Example" box). As you will have guessed, you need to alternately add 3 and 2 to the existing numbers in the 2,5,7,10,12 series. Since the jump from 10 to 12 was two units in length, a count of three thus needs to be added to predict the next number: 15 is the correct result. The network tended toward 14 most of the time, so it cannot compete with human intelligence (as of yet).

Figure 8: The LSTM network did not do well in the intelligence test, because it failed to identify the different growth rates.