Contextual Memory Trees

TL;DR

The paper introduces a Contextual Memory Tree (CMT), a scalable memory controller that efficiently manages and retrieves unbounded experience data, enhancing existing learning algorithms with improved performance and computational efficiency.

Contribution

It presents a novel memory structure that supports efficient insertion and retrieval, integrates seamlessly with learning algorithms, and demonstrates improved classification and image-captioning performance.

Findings

CMT achieves logarithmic time insertion and retrieval.

Augmenting classifiers with CMT improves statistical performance.

CMT outperforms simple nearest neighbors in image-captioning tasks.

Abstract

We design and study a Contextual Memory Tree (CMT), a learning memory controller that inserts new memories into an experience store of unbounded size. It is designed to efficiently query for memories from that store, supporting logarithmic time insertion and retrieval operations. Hence CMT can be integrated into existing statistical learning algorithms as an augmented memory unit without substantially increasing training and inference computation. Furthermore CMT operates as a reduction to classification, allowing it to benefit from advances in representation or architecture. We demonstrate the efficacy of CMT by augmenting existing multi-class and multi-label classification algorithms with CMT and observe statistical improvement. We also test CMT learning on several image-captioning tasks to demonstrate that it performs computationally better than a simple nearest neighbors memory system while benefitting from reward learning.