From a single decision to a multi-step algorithm

Humans can perform sequential and recursive computations, as when calculating 23 x 74. However, this comes at a cost: flexible computations are slow and effortful. We argue that this competence involves serial chains of successive decisions, each based on the accumulation of evidence up to a threshold and forwarding the result to the subsequent step. Such serial 'programs' require a specific neurobiological architecture, approximating the operation of a slow serial Turing machine. We review recent progress in understanding how the brain implements such multi-step decisions and briefly examine how they might be realized in models of primate cortex.