Neural Networks have gone from academic curiosity to reviving hopes for Artificial Intelligence.
It seems like every day brings another break through application.
One part of this story is scale, another is design.
Neural Networks have been able to bring machine learning to non-tabular datasets because they engineer their own features.
While any data can be put into a number grid, tabular data have features with a consistent relationship to the outcome across observations.
That allows machine learning to proceed by identifying that consistent relationship, however complex, noisy or weak it might be.
Non-tabular data doesn’t simply have consistently informative raw features; patterns with a predictive value of the outcome can appear in different locations within observations. Informative features have to be constructed and the right network architectures encourage the training of features that learn invariances that are informative of the outcome a model is trying to predict and exclude uninformative variances that occur within a pattern.
At DLG we have established a deep learning development program that can effectively identify high risk driving from telematic driving data.
Technical level: High level/overview
Supported by
