For this purpose I have trained a convolutional neural network with TensorFlow, under python 3.7. The network consisted of 9 layers in total, from input to output: an Average Pooling input layer (6 × 3 pooling size); 3 convolutional layers (3 × 3 kernel, ‘relu’ activation function, with 64, 16, 32 neurons, respectively); 3 pooling layers (MaxPool 2 × 2; Dropout; Flatten); 1 Dense layer (with a ‘relu’ activation function); and a final Dense output layer (with a ‘softmax’ activation function). The network was compiled with an ‘adam’ optimizer, a sparse categorical ‘crossentropy’ function, and ‘accuracy’ as metrics. The fitting was done in 10 epochs and with a batch size of 32. The output layer, representing the animals in each group, contained an additional neuron, here called null, that was trained on pictures triggered by the animals by accident (e.g., with their back).

The same approach was used with rhesus macaques for a study that is at the moment still in preparation. Overall, between long-tailed and rhesus macaques, this approach was used to identify 11 animals, belonging to 4 different housing groups. As shown in the figure below, while the number of animals for each network made no difference for the network accuracy (MultiNet vs OneNet), the framework used for training the network did (Keras vs CoreML). The model I designed and trained my self performed significantly better than the black box model trained with CoreML on a MacBook Pro (measured across 8 animals and over 30k data points).

In order to find the appropriate network configuration for the task, I have automated a bootstrap procedure to find the best parameters and run it on Google Colab to make use of GPU computations. Parameters include the size of the average pooling kernel (AvPool); the number of units in the three convolutional hidden layers (Conv_1, Conv_2, Conv_3); and the number of units in the hidden dense layer (DenLay). With a test dataset of 3,000 pictures of two male rhesus macaque taken with the same device and in the same facility, we trained and tested 46 combinations of the parameters mentioned before. Finally, we compared the performances of the 46 resulting networks and handpicked the combination of parameters of the network with the highest accuracy (98.7%). This combination was used as the final configuration for the network used during the experiment.