[論文レビュー] Deep Generative Dual Memory Network for Continual Learning

Nitin Kamra, Umang Gupta+1 more|arXiv (Cornell University)|Oct 28, 2017

Domain Adaptation and Few-Shot Learning参考文献 31被引用数 124

ひとこと要約

双重メモリアーキテクチャと深部生成リプレイを提案し、逐次タスク学習における破局的忘却を緩和。複数の画像分類ベンチマークで保持能力の改善を示す。

ABSTRACT

Despite advances in deep learning, neural networks can only learn multiple tasks when trained on them jointly. When tasks arrive sequentially, they lose performance on previously learnt tasks. This phenomenon called catastrophic forgetting is a fundamental challenge to overcome before neural networks can learn continually from incoming data. In this work, we derive inspiration from human memory to develop an architecture capable of learning continuously from sequentially incoming tasks, while averting catastrophic forgetting. Specifically, our contributions are: (i) a dual memory architecture emulating the complementary learning systems (hippocampus and the neocortex) in the human brain, (ii) memory consolidation via generative replay of past experiences, (iii) demonstrating advantages of generative replay and dual memories via experiments, and (iv) improved performance retention on challenging tasks even for low capacity models. Our architecture displays many characteristics of the mammalian memory and provides insights on the connection between sleep and learning.

研究の動機と目的

• Sequential task learning における破局的忘却を克服する動機。
• 海馬と新皮質に触発された dual memory アーキテクチャを導入。
• 過去の知識を長期記憶に統合するための生成リプレイを利用。
• デュアルメモリと生成リプレイがベースラインを上回ることを示す（低容量モデルを含む）。
• 睡眠のような統合と継続的学習との関連性を強調。

提案手法

• Generator G、learner L、dictionary D_dgm を備えた Deep Generative Memory (DGM) を導入。
• Deep Generative Replay (DGR) を用いて新規データと過去タスクからの生成サンプルの混合で訓練。
• デュアルメモリシステムを実装：新規タスク用の複数の STTM ユニットを持つ Short-Term Memory (STM) と知識を統合する Long-Term Memory (LTM)。
• 睡眠中、STM がサンプルを生成し、生成リプレイを介して LTM に統合。
• 生成器として Variational Autoencoder (VAE) を用い、リプレイ時にサンプルを再構成・ノイズ除去。
• Sequential task benchmarks で評価し、NN、Dropout、PPR、EWC、DGR などのベースラインと比較。

実験結果

リサーチクエスチョン

• RQ1 Can a dual memory architecture paired with generative replay prevent catastrophic forgetting during sequential task learning?
• RQ2 How does the proposed method compare to existing baselines in terms of average accuracy (ACC) and backward transfer (BWT)?
• RQ3 What is the impact of memory constraints and task revision on continual learning performance?
• RQ4 How does sleep-like consolidation (periodic memory integration) influence long-term retention and training efficiency?

主な発見

Algorithm	Digits_ACC	Permnist_ACC	Shapes_ACC	Hindi_ACC
NN	0.1	0.588	0.167	0.125
DropNN	0.1	0.590	0.167	0.125
PPR	0.1	0.574	0.167	0.134
EWC	0.1	0.758	0.167	0.125
DGR	0.596	0.861	0.661	0.731
DGDMN	0.818	0.831	0.722	0.658

• DGDMN and DGR consistently outperform baselines on average accuracy across multiple datasets.
• DGDMN achieves higher ACC than DGR on several datasets (Digits, Permnist, Shapes, Hindi).
• DGDMN attains the least negative backward transfer, indicating reduced forgetting compared to baselines.
• With memory constraints, DGDMN remains more robust and trains faster than DGR, especially over long task sequences.
• The dual-memory + generative replay approach demonstrates gradual forgetting rather than catastrophic forgetting in long sequences.