Latent Space Exploration and Functionalization of a Gated Working Memory Model Using Conceptors

Working memory is the ability to maintain and manipulate information. We introduce a method based on conceptors that allows us to manipulate information stored in the dynamics (latent space) of a gated working memory model. This latter model is based on a reservoir: a random recurrent network with trainable readouts. It is trained to hold a value in memory given an input stream when a gate signal is on and to maintain this information when the gate is off. The memorized information results in complex dynamics inside the reservoir that can be faithfully captured by a conceptor. Such conceptors allow us to explicitly manipulate this information in order to perform various, but not arbitrary, operations. In this work, we show (1) how working memory can be stabilized or discretized using such conceptors, (2) how such conceptors can be linearly combined to form new memories, and (3) how these conceptors can be extended to a functional role. These preliminary results suggest that conceptors can be used to manipulate the latent space of the working memory even though several results we introduce are not as intuitive as one would expect.