T.A. de Lorm, C. Horswill, D. Rabaiotti, R.M. Ewers, R. J.
Groom, J. Watermeyer, R. Woodroffe
Table S1 The structure of a Convolutional Neural Net that
classifies images of African Wild Dogs into “standing” or “not
standing” (S1.A), and a Convolutional Neural Net that classifies images
of African Wild Dogs into “left flank” or “right flank” (S1.B). The
layers are given in the order at which they occur in the model. The
models were optimised using RMSprop, an algorithm which guides how the
model improves itself (Tieleman & Hinton, 2012). The activation column
refers to which activation function was used in each layer, which
determines how the nodes within layers convert its input to an
output-value. ReLu was used as activator function, as this has been
found to improve multi-layer networks (Glorot, Bordes & Bengio, 2011).
The last layer is activated with a Sigmoid function, which turns the
input into a single, binary prediction (“standing” or “not
standing”, “left” or “right”).