Learning and analyzing vector encoding of symbolic representations
This work addresses the challenge of representing and querying symbolic data in neural networks, but it appears incremental as it builds on existing theoretical methods.
The authors tackled the problem of learning vector encodings for symbolic structures using a formal language and a sequence-to-sequence network, achieving an approximate linearity property similar to theoretical techniques.
We present a formal language with expressions denoting general symbol structures and queries which access information in those structures. A sequence-to-sequence network processing this language learns to encode symbol structures and query them. The learned representation (approximately) shares a simple linearity property with theoretical techniques for performing this task.