We live in highly structured environments where patterns often repeat. Our brain forms internal models of the world by extracting the regularities. Such models are utilised to form predictions about future events and incoming information. Predictions lead to efficient neural processing of incoming stimuli and hence facilitate perception. Although the existence of prediction signals is known, much less is known about the neural mechanisms underlying the propagation of prediction signals, especially in higher-vision. Predictive processing theories propose that in cortical hierarchies, high-level prediction signals are sent to lower regions via feedback pathways, where they are compared against its input to compute a prediction error. In this study we investigate neural mechanisms underlying communication of sensory predictions using functional neuroimaging. To this end, we leverage the macaque face-processing network, a three-level hierarchical system in the ventral-visual pathway where face representation becomes more view-invariant as information goes up the hierarchy. We test the role of feedback pathways in sending predictions by investigating how expectations affect neural representations. We hypothesized that expectations lead to higher-order area sending predictions, such that the lower-areas inherit the tuning properties of the areas from which they receive feedback. By conducting representational similarity analyses, we show that after statistical learning of arbitrary face-pair sequences, expectations lead to view-invariant representations in lower face-areas. Rather than their own view-specific feedforward tuning properties, these lower areas now exhibit view-invariant abstract representations of higher face-areas. This cascading-down of high-level prediction signals in the entire face-processing network suggests a functional role of feedback connections in signaling predictions, which is in-line with predictive processing theories. By showing how the top-down information flow of predictions and previous experience affects face-processing, this work contributes to a revision of currently dominant theories that view face perception and generally, object recognition through the lens of pure feedforward architectures.