The Neural Executive: Can Gated Attractor Networks Account for Cognitive Control?

AI systems can be grouped into two main categories: symbolic and neural. Each has advantages in certain types of tasks, and to date symbolic systems are typically superior at cognitive control and executive function. This paper presents a model using the GALIS framework which probes at this gap by successfully completing the n-Back cognitive control task at human-like levels using a neural approach. Several aspects of the model discussed here, in particular gated sequential attractor networks and behavior driven in part from the learned contents of instruction memory, allow it to behave in ways that are more typical of symbolic systems.