[논문 리뷰] Pseudo-Recursal: Solving the Catastrophic Forgetting Problem in Deep Neural Networks

Craig Atkinson, Brendan McCane외 2명|arXiv (Cornell University)|2018. 02. 12.

Domain Adaptation and Few-Shot Learning참고 문헌 13인용 수 54

한 줄 요약

본 논문은 pseudo-recursal을 도입하여 pseudo-rehearsal과 GAN이 생성한 pseudo-items를 결합해 CIFAR-10, SVHN, MNIST에서 과제별 메모리 확장 없이 심층 네트워크의 지속적 학습(연속 학습)을 달성하고, 망각을 크게 감소시킨다.

ABSTRACT

In general, neural networks are not currently capable of learning tasks in a sequential fashion. When a novel, unrelated task is learnt by a neural network, it substantially forgets how to solve previously learnt tasks. One of the original solutions to this problem is pseudo-rehearsal, which involves learning the new task while rehearsing generated items representative of the previous task/s. This is very effective for simple tasks. However, pseudo-rehearsal has not yet been successfully applied to very complex tasks because in these tasks it is difficult to generate representative items. We accomplish pseudo-rehearsal by using a Generative Adversarial Network to generate items so that our deep network can learn to sequentially classify the CIFAR-10, SVHN and MNIST datasets. After training on all tasks, our network loses only 1.67% absolute accuracy on CIFAR-10 and gains 0.24% absolute accuracy on SVHN. Our model's performance is a substantial improvement compared to the current state of the art solution.

연구 동기 및 목표

• Addresses catastrophic forgetting in sequential task learning for DNNs.
• Propose a memory-efficient continual learning approach that does not grow with tasks.
• Leverage Generative Adversarial Networks to generate representative pseudo-items for rehearsal.
• Demonstrate recursive pseudo-rehearsal (pseudo-recursal) applied to both classifier and generator.
• Compare against Elastic Weight Consolidation and standard pseudo-rehearsal on image datasets.

제안 방법

• Formalize pseudo-rehearsal for sequential tasks with a fixed architecture.
• Use a GAN to generate pseudo-images representing past tasks for rehearsal.
• Apply pseudo-rehearsal recursively to the GAN to cover multiple tasks without extra memory per task.
• Train classifier and GAN on current task plus pseudo-items representing previous tasks.
• Evaluate on CIFAR-10, SVHN, and MNIST with multiple experimental conditions (std, reh, pseudo_rec, ewc, ewc_c10, rote_learn).
• Measure accuracy retention on previous tasks after learning new tasks; report absolute accuracy changes.

실험 결과

연구 질문

• RQ1 Can pseudo-rehearsal with GAN-generated pseudo-images prevent catastrophic forgetting across CIFAR-10, SVHN, and MNIST without increasing memory per task?
• RQ2 Does applying pseudo-rehearsal recursively to both classifier and GAN yield better retention than standard pseudo-rehearsal or EWC?
• RQ3 How does pseudo-recursal perform relative to baselines in terms of retaining previous task performance after sequential learning?
• RQ4 What are the trade-offs in training time and memory when using a GAN-based pseudo-rehearsal approach?

주요 결과

• The method retains CIFAR-10 accuracy with only a 1.67% absolute drop after learning all tasks.
• SVHN accuracy increases by an absolute 0.24% after training on subsequent tasks.
• Pseudo-recursal outperforms EWC in retaining prior task performance across CIFAR-10 and SVHN.
• Compared to rote learning baselines, pseudo-recursal yields substantial maintenance of past task accuracy (e.g., 9.6% to 13.11% gains on CIFAR-10 and SVHN, respectively).
• Recursively training the GAN on past tasks allows generating representative pseudo-items without extra memory per task.
• The approach demonstrates effective continual learning without hard constraints on intermediate neurons and without storing past task data.