Neural Information Processing Systems http://nips.cc/

Most recently proposed methods for Neural Program induction work under the assumption of having a large set of input/output (I/O) examples for learning any given input-output mapping. This paper aims to address the problem of data and computation efficiency of program induction by leveraging information from related tasks. Specifically, we propose two novel approaches for cross-task knowledge transfer to improve program induction in limited-data scenarios. In our first proposal, portfolio adaptation, a set of induction models is pretrained on a set of related tasks, and the best model is adapted towards the new task using transfer learning. In our second approach, meta program induction, a $k$-shot learning approach is used to make a model generalize to new tasks without additional training.

Neural Information Processing Systems http://nips.cc/

Much of everyday human thinking and learning can be understood in terms of program induction: constructing a procedure that maps inputs to desired outputs, based on observing example input-output pairs.

This work extends the ProbLog language and uses the distribution of grounded facts estimated by the ProbLog to train neural networks, which is represented as neural predicates in the ProbLog. Meanwhile, the DeepProbLog framework is able to learn ProbLog parameters and deep neural networks at the same time. The experimental results show that the DeepProbLog can perform joint probabilistic logical reasoning and neural network inference on some simple tasks. Combining perception and symbolic reasoning is an important challenge for AI and machine learning. Different to most of the existing works, this work does not make one side subsumes the other (e.g.

The paper is making an evaluation of several approaches for neural network-based induction of computer programs. The main proposal is on the use of meta-learning, in order to exploit knowledge learned from various tasks, for learning on a specific task with a small number of instances. For that purpose, three approaches are proposed: 1) transfer learning for adapting an existing model trained on a related task; 2) meta program induction, where the model has been trained to work on a variety of tasks; and 3) meta program adapted for a given task. The paper also proposes to make use of a synthetic domain Karel, which comes from an educational language for teach computer programming, which consists in moving a robot in a 2D grid through computer instructions. Results are reported with varying the number of instances used for program induction for the three meta approaches proposed plain method, with results showing some advantages with little number of instances.