Improving Neural Turing Machine and applying it to human behaviour pattern prediction
- Kordík Pavel
Our recent research article which was accepted for publication in the proceedings of World Congress of Computational Intelligence (WCCI 2016) presents our experiments with Neural Turing Machine (NTM), recently proposed by Google researchers.
We published one of the first NTM open source implementation which was able to repeat experiments in the paper.
Recently, we work on improvements that enable faster and more stable NTM learning.
NTM proved that it is very powerful in learning and generalizing long sequences. It can outperform standard recurrent neural networks as well as popular gating recurrent nets (LSTM).
We extended NTM to be able to efficiently predict sequential patterns. First, we clustered GPS data.
And then we trained NTM on historical data (sequence of cluster ids) to predict how people move along clusters.
We cooperate with Czech police and applied this method to model behavior of suspects.