Figure 2
(a) Representation of self-attention. Connections are maintained throughout sequence as model trains. The purple circles indicate the predictions, and the dotted arrows are the attention mechanism keeping all information connected between predictions so that downline future predictions still retain and have access to the information in the earlier predictions. The model predicts × 2′ for the first input of data, uses the predicted data to predict the next value × 3′. After all predictions are made, the loss is calculated between actual input (× 1… × 5) and predicted outputs (× 2′… × 6′). (b) The input is the malaria timeseries and climate data, the target is the sequence shifted to the right by one time step so for each new input, the model will output a prediction.
