When learning an input-output mapping from very few examples, is it better to first infer a latent function that explains the examples, or is it better to directly predict new test outputs, e.g. using a neural network? We study this question on ARC by training neural models for \emph{induction} (inferring latent functions) and \emph{transduction} (directly predicting the test output for a given test input). We train on synthetically generated variations of Python programs that solve ARC training tasks. We find inductive and transductive models solve different kinds of test problems, despite having the same training problems and sharing the same neural architecture: Inductive program synthesis excels at precise computations, and at composing multiple concepts, while transduction succeeds on fuzzier perceptual concepts. Ensembling them approaches human-level performance on ARC.

We have created a hybrid neuro-robotic system that establishes two-way communication between the brain of a lamprey and a small mobile robot. The purpose of this system is to offer a new paradigm for investigating the behavioral, computational, and neurobiological mechanisms of sensory-motor learning in a unified context. The mobile robot acts as an artificial body that delivers sensory information to the neural tissue and receives command signals from it. The sensory information encodes the intensity of light generated by a fixed source. The closed-loop interaction between brain and robot generates autonomous behaviors whose features are strictly related to the structure and operation of the neural preparation. We provide a detailed description of the hybrid system, and we present experimental findings on its performance. In particular, we found (a) that the hybrid system generates stable behaviors, (b) that different preparations display different but systematic responses to the presentation of an optical stimulus, and (c) that alteration of the sensory input leads to short- and long-term adaptive changes in the robot responses. The comparison of the behaviors generated by the lamprey's brain stem with the behaviors generated by network models of the same neural system provides us with a new tool for investigating the computational properties of synaptic plasticity.