Fig. 7: The network is capable of forecasting multiple movies without being retrained.
From: Waves traveling over a map of visual space can ignite short-term predictions of sensory input
a The recurrent network model was adapted to contain a higher-level competitive-learning process. Left: Readout matrices were learned separately for separate examples. Right: Storing the learned readout matrices in an aggregate matrix \({{\mathcal{V}}}\), the present network state drove the aggregate matrix toward either of the learned matrices via an unsupervised competitive learning rule. b Beginning with feeding frames from movie 1, the network takes some time to recall the learned matrix that results in an accurate closed-loop forecast. Quickly switching to a different movie, the network once again takes some time to adjust its output weights before converging to the correct ones for an accurate closed-loop forecast.
