[논문 리뷰] Learning to Remember Rare Events

Łukasz Kaiser, Ofir Nachum외 2명|arXiv (Cornell University)|2017. 03. 09.

Domain Adaptation and Few-Shot Learning참고 문헌 20인용 수 237

한 줄 요약

학습 가능한 키-값 메모리에서 학습된 저장소를 통한 빠른 최근접 이웃 검색으로 장기 지속 학습(One-shot) 가능하게 하는 확장 가능한 평생 기억 모듈을 신경망에 도입하여 Omniglot에서 최첨단 성능을 달성하고 메모리 기반 한 샷 기능을 통해 번역을 향상시킵니다.

ABSTRACT

Despite recent advances, memory-augmented deep neural networks are still limited when it comes to life-long and one-shot learning, especially in remembering rare events. We present a large-scale life-long memory module for use in deep learning. The module exploits fast nearest-neighbor algorithms for efficiency and thus scales to large memory sizes. Except for the nearest-neighbor query, the module is fully differentiable and trained end-to-end with no extra supervision. It operates in a life-long manner, i.e., without the need to reset it during training. Our memory module can be easily added to any part of a supervised neural network. To show its versatility we add it to a number of networks, from simple convolutional ones tested on image classification to deep sequence-to-sequence and recurrent-convolutional models. In all cases, the enhanced network gains the ability to remember and do life-long one-shot learning. Our module remembers training examples shown many thousands of steps in the past and it can successfully generalize from them. We set new state-of-the-art for one-shot learning on the Omniglot dataset and demonstrate, for the first time, life-long one-shot learning in recurrent neural networks on a large-scale machine translation task.

연구 동기 및 목표

• 평생 설정에서 희귀 이벤트로부터 학습하는 문제를 동기 부여하고 해결한다.
• 훈련 중 업데이트되는 키-값 쌍을 저장하는 미분 가능한 메모리 모듈을 제안한다.
• 추론 시 메모리 키에 대한 최근접 이웃 검색을 이용하여 원샷 학습을 가능하게 한다.
• 모듈을 CNN, Seq2Seq, GNMT에 통합하고 Omniglot, 합성 과제, 번역에서의 활용성을 평가하여 다재다능함을 보여준다.]
• method:[

제안 방법

• Memory module stores keys K, values V, and age A as a memory M of size memory-size.
• Query q (normalized) retrieves k=256 nearest neighbors via cosine similarity, returning V of the top neighbor and a softmax-weighted similarity signal.
• Memory loss uses a margin-based triplet objective comparing positive and negative neighbors, encouraging proximity of q to the correct key and separation from incorrect ones.
• Memory updates: if the retrieved value matches target v, update the key by averaging with q; otherwise write (q,v) to the oldest memory slot (with small random perturbation).
• Efficient NN: exact computation via QK^T or approximate via locality-sensitive hashing (LSH) for large memory.
• Applied across architectures: simple CNN, GNMT-style seq2seq, and Extended Neural GPU to demonstrate broad compatibility.

실험 결과

연구 질문

• RQ1 Can a differentiable, scalable memory module enable life-long one-shot learning across diverse neural architectures?
• RQ2 Does integrating memory improve performance on standard one-shot tasks (Omniglot) and synthetic life-long tasks, and can it aid large-scale translation?
• RQ3 How does memory influence learning and generalization when rare events or words appear?
• RQ4 What are the practical effects and metrics for evaluating one-shot, lifelong learning in translation and other sequence tasks?

주요 결과

• Memory-augmented models achieve strong one-shot learning on Omniglot, approaching or matching state-of-the-art results.
• On a synthetic task designed to require memory, memory-augmented models significantly outperform baselines and standard seq2seq models.
• In GNMT English–German translation, memory-augmented models perform on par with baseline BLEU scores and show one-shot gains when context memory is used; exposing the whole test set as memory context yields substantial BLEU improvement (8+ points).
• Qualitative example shows the memory module translates rare words like Dostoevsky, which baseline models struggle to translate.
• Across architectures and tasks, a single set of memory parameters (k=256, α=0.1) yields good results, illustrating versatility.