When comparing human with artificial intelligence, one major difference is\napparent: Humans can generalize very broadly from sparse data sets because they\nare able to recombine and reintegrate data components in compositional manners.\nTo investigate differences in efficient learning, Joshua B. Tenenbaum and\ncolleagues developed the character challenge: First an algorithm is trained in\ngenerating handwritten characters. In a next step, one version of a new type of\ncharacter is presented. An efficient learning algorithm is expected to be able\nto re-generate this new character, to identify similar versions of this\ncharacter, to generate new variants of it, and to create completely new\ncharacter types. In the past, the character challenge was only met by complex\nalgorithms that were provided with stochastic primitives. Here, we tackle the\nchallenge without providing primitives. We apply a minimal recurrent neural\nnetwork (RNN) model with one feedforward layer and one LSTM layer and train it\nto generate sequential handwritten character trajectories from one-hot encoded\ninputs. To manage the re-generation of untrained characters, when presented\nwith only one example of them, we introduce a one-shot inference mechanism: the\ngradient signal is backpropagated to the feedforward layer weights only,\nleaving the LSTM layer untouched. We show that our model is able to meet the\ncharacter challenge by recombining previously learned dynamic substructures,\nwhich are visible in the hidden LSTM states. Making use of the compositional\nabilities of RNNs in this way might be an important step towards bridging the\ngap between human and artificial intelligence.\n