[Paper Review] An Empirical Investigation of Catastrophic Forgetting in Gradient-Based Neural Networks

Ian Goodfellow, Mehdi Mirza+3 more|arXiv (Cornell University)|Jan 1, 2014

Domain Adaptation and Few-Shot Learning14 references497 citations

TL;DR

This paper empirically investigates catastrophic forgetting in gradient-based neural networks across various architectures, activation functions, and task relationships. It finds that dropout consistently outperforms other training methods in balancing retention of prior knowledge and adaptation to new tasks, while activation function performance is highly task-dependent, necessitating cross-validation for optimal selection.

ABSTRACT

Abstract: Catastrophic forgetting is a problem faced by many machine learning models and algorithms. When trained on one task, then trained on a second task, many machine learning models forget how to perform the first task. This is widely believed to be a serious problem for neural networks. Here, we investigate the extent to which the catastrophic forgetting problem occurs for modern neural networks, comparing both established and recent gradient-based training algorithms and activation functions. We also examine the effect of the relationship between the first task and the second task on catastrophic forgetting. We find that it is always best to train using the dropout algorithm--the dropout algorithm is consistently best at adapting to the new task, remembering the old task, and has the best tradeoff curve between these two extremes. We find that different tasks and relationships between tasks result in very different rankings of activation function performance. This suggests the choice of activation function should always be cross-validated.

Motivation & Objective

• To assess the extent of catastrophic forgetting in modern neural networks during sequential task learning.
• To compare the performance of established and recent gradient-based training algorithms in mitigating forgetting.
• To evaluate how the relationship between tasks influences forgetting behavior.
• To determine the impact of different activation functions on retention and adaptation.
• To identify optimal training strategies that balance remembering old tasks and learning new ones.

Proposed method

• Conducted empirical experiments on neural networks trained sequentially on multiple tasks using various gradient-based optimization algorithms.
• Evaluated performance across multiple activation functions, including ReLU, ELU, and others, under different task relationships.
• Measured catastrophic forgetting using accuracy on the first task after training on the second task.
• Used dropout regularization as a baseline and comparison method across all experiments.
• Analyzed trade-offs between retention of old task performance and adaptation to new tasks via performance curves.
• Systematically varied the similarity and overlap between first and second tasks to assess their influence on forgetting.

Experimental results

Research questions

• RQ1 To what extent does catastrophic forgetting occur in modern neural networks during sequential training?
• RQ2 How do different gradient-based training algorithms compare in mitigating forgetting?
• RQ3 How does the relationship between the first and second task affect the degree of forgetting?
• RQ4 Which activation function yields the best balance between remembering old tasks and learning new ones?
• RQ5 Is there a universally optimal training configuration that minimizes forgetting across diverse task pairs?

Key findings

• Dropout consistently outperformed all other training algorithms in preserving performance on the first task while adapting to the second.
• The performance ranking of activation functions varied significantly depending on the relationship between the first and second tasks.
• No single activation function was universally optimal; performance was highly task-dependent, necessitating cross-validation.
• The trade-off between remembering old tasks and learning new ones was most favorable when using dropout.
• Task similarity and overlap had a strong influence on forgetting rates, with more similar tasks showing less forgetting.
• Even among modern architectures, catastrophic forgetting remains a significant issue unless mitigated by appropriate training strategies like dropout.