REPRISE: A Retrospective and Prospective Inference Scheme
- Martin Butz, Department of Computer Science and Department of Psychology / Faculty of Science / Cognitive Modeling, University of Tübingen, Tübingen, Germany
- David Bilkey, Department of Psychology, University of Otago, Dunedin, New Zealand
- Alistair Knott, Computer Science, University of Otago, Dunedin, New Zealand
- Sebastian Otte, Department of Computer Science / Faculty of Science / Cognitive Modeling, University of Tübingen, Tübingen, Germany
AbstractMotivated by the close relation of predictive coding and active inference to cognition, we introduce a dynamic artificial neural network-based (ANN) adaptation process, which we term REPRISE: REtrospective and PRospective Inference SchEme. REPRISE first executes a retrospective inference process, inferring the unobservable contextual state that best explains its recently encountered sensorimotor experiences. It then executes a prospective inference process, inferring upcoming motor activities in the light of the inferred contextual state and a given goal state. First, the ANN – a recurrent neural network – is trained to learn one sensorimotor temporal forward model, that is, the sensorimotor contingencies generated by the behavior of three moving or flying vehicles. During training, additional three bits are provided as input, indicating which mode currently applies. After training, goal-directed control and system state inference are activated: Given a goal state, the system imagines a motor command sequence optimizing it with the prospective objective to minimize the distance to the goal. Meanwhile, the system evaluates the encountered sensorimotor contingencies retrospectively, adapting its vehicle estimation activities and, in order to maintain coherence, the neural hidden states accordingly. This ANN’s ’mind’ is thus continuously imagining the future and reflecting on the past – showing superior performance on the posed control problems. The architecture effectively demonstrates that neural error signals and neural activities can be projected into the past and into the future, respectively, optimizing both neural context codes that approximately generate the recent past and upcoming behavior in the light of desired goal states.
Return to previous page