A fundamental step in using machine learning to model some system or process is the training of the network. This involves finding the network parameter settings that produce the best predictions of some training data, commonly achieved by using gradient descent to minimise the difference between prediction and training data. PolyNet uses the PolyChord engine to fully sample all possible parameter settings of a network, allowing one to leverage the full predictive power of a given network architecture, rather than just using a single or small handful of predictions.
PolyChord Algorithm Parsing Rastrigin Fuction
This full sampling makes for more efficient networks since a smaller network that leverages all its predictions will do substantially better than one that does not. It also provides uncertainties on the predictions, which makes the network explainable and less of a black box. Furthermore, the resulting predictions are far more generalisable, which combines with the uncertainty to mitigate the risk of false positives/negatives, providing unprecedented trust in PolyNet's predictions.
Importantly, using the model-selection power of the PolyChord technology, PolyNet can be used in network design by allowing the user to rate the performance of different types of network. To learn more, head to our contact us page and shoot us a message for more details.