Encoding Longer-Term Contextual Information with Predictive Coding and Ego-Motion

Figure 1

A 2-layer MT-AFA-PredNet. A 2-layer MT-AFA-PredNet Example: the green (generative) units calculate (Eq. (9)) for prediction and further combine proportionally by the attention unit (Eq. (10)). The red units calculate the error (Eq. (8)). The error is obtained from blue units (Eq. (7) (Prediction) and Eq. (6) (Target)). Moreover, the time constant changes on different levels.