Differentiable Programs with Neural Libraries

TL;DR

This paper introduces a framework combining differentiable programming with neural networks to create interpretable, modular algorithms that improve through lifelong learning and transfer knowledge across tasks.

Contribution

It presents a novel framework for differentiable programs with neural libraries, enabling end-to-end training and modularity for better generalization and lifelong learning.

Findings

Modular neural libraries enhance transfer learning.

Framework achieves better lifelong learning performance.

Models learn interpretable algorithms with perceptual components.

Abstract

We develop a framework for combining differentiable programming languages with neural networks. Using this framework we create end-to-end trainable systems that learn to write interpretable algorithms with perceptual components. We explore the benefits of inductive biases for strong generalization and modularity that come from the program-like structure of our models. In particular, modularity allows us to learn a library of (neural) functions which grows and improves as more tasks are solved. Empirically, we show that this leads to lifelong learning systems that transfer knowledge to new tasks more effectively than baselines.