Inferring Event-Predictive Goal-Directed Object Manipulations in REPRISE

The recently introduced REtrospective and PRospective Inference SchEme (REPRISE) infers contextual event states in the form of neural parametric biases retrospectively in recurrent neural networks (RNNs), distinguishing, for example, different sensorimotor control dynamics. Moreover, it actively infers motor commands prospectively in a goal-directed manner, minimizing anticipated future loss signals—such as the distance to a goal location. REPRISE struggles, however, when multiple, somewhat competing goals are active in parallel—such as when an object is to-be picked up and carried to a goal location. Moreover, unsuitable statistical correlations in the training data can prevent successful goal-directed motor inference, failing to reach particular goal constellations. We scrutinize this challenge and propose that appropriate gradient separation techniques are missing. First, we show that relative encodings and suitable training schedules can alleviate the problem. Most robust behavior, however, is achieved when the RNN architecture is suitably modularized. In the future, emergent RNN modularizations and more direct gradient separation mechanisms need to be developed. Moreover, we expect that REPRISE will shed further light onto the hierarchical neuro-cognitive structure of human thought.