Neural Latching Switch Circuits for implementing Finite-State automata

Cognitive processes rely on switches between mental states. Mental states are thought to be supported by cell assembly activations as introduced by Hebb. This idea has been formalized using attractor neural networks, leading to detailed mechanistic and quantitative characterizations, allowing for confrontations with neurobiology experiments. However such a mechanistic understanding is lacking for switches between mental states. Here we introduce Neural Latching Switch Circuits (NLSC) which are composed of an attractor neural network, augmented with gate neurons allowing to program transitions between network states. We explore conditions under which such circuits emerge in artificial neural networks, and provide a quantitative description by embodying NLSC into models of binary neurons. We show how NLSC can be mapped onto the flys head-direction system and put forward signatures of NLSC for identifying such a structure in other brain circuits. Throughout examples, we show that NLSC are suited to implement computations relevant for sensory, motor, or more abstract types of processing. To propose a meaningful characterization of NLSC in these various contexts, we interpret them as finite-state automata, identifying attractor-states and automata-states, a first step in establishing a connection between neural and symbolic computations.