Algorithmic learning, a next step for AI. An application to arithmetic operations
Abstract
Deep learning achieved state of the art performances in multiple domains (image recognition, natural language processing, etc.) One of the next steps is to be able to learn algorithms, as a way to provide some new forms of generalization for AI systems. This is currently a hard and challenging problem as it involves algorithmic recurrence, memory management and combination of subtasks, which leads to trainability problems. We show that even a simple algorithm of multiplication manifests trainability problems for neural networks.
In this article, we present an original training method applied to multi-digit multiplication learning, called Unrolling Algorithmic Training (UAT).
To learn the global algorithm, we use additional supporting tasks consisting in the successive subtasks composing the global algorithm (in our case the 1-digit multiplications and the final addition). The global end-to-end and the subtasks learning are then balanced with an active learning mechanism.
This multi-task learning allows to overcome the problem of trainability encountered when learning directly the global algorithmic task. Our interpretation is that the network is able to somehow combine the subtasks in order to learn the global task.
Moreover, we show that the global algorithm can be bootstrapped, fine-tuned and is even resilient without retraining it from scratch when we vary the size of recurrence provided to the network.
Origin | Files produced by the author(s) |
---|