Evolving Causal Neural Networks

We introduce causal neural networks, a generaliza-tion of the usual feedforward neural networks which allows input features and target outputs to be rep-resented as input or output units. For inferring the values of target outputs which are represented as in-put units, we developed a forward-backward propa-gation algorithm which uses gradient descent to min-imize the error of the predicted output features. To deal with the large number of possible structures and feature selection, we use a genetic algorithm. Exper-iments on a regression problem and 5 classication problems show that the causal neural networks can outperform the usual feedforward architectures for particular problems.