Measuring Catastrophic Forgetting in Neural Networks

Abstract: Deep neural networks are used in many state-of-the-art systems for machine perception. Once a network is trained to do a specific task, e.g., bird classification, it cannot easily be trained to do new tasks, e.g., incrementally learning to recognize additional bird species or learning an entirely different task such as flower recognition. When new tasks are added, typical deep neural networks are prone to catastrophically forgetting previous tasks. Networks that are capable of assimilating new information incrementally, much like how humans form new memories over time, will be more efficient than re-training the model from scratch each time a new task needs to be learned. There have been multiple attempts to develop schemes that mitigate catastrophic forgetting, but these methods have not been directly compared, the tests used to evaluate them vary considerably, and these methods have only been evaluated on small-scale problems (e.g., MNIST). In this paper, we introduce new metrics and benchmarks for directly comparing five different mechanisms designed to mitigate catastrophic forgetting in neural networks: regularization, ensembling, rehearsal, dual-memory, and sparse-coding. Our experiments on real-world images and sounds show that the mechanism(s) that are critical for optimal performance vary based on the incremental training paradigm and type of data being used, but they all demonstrate that the catastrophic forgetting problem has yet to be solved.

• The paper introduces novel metrics (Ω_base, Ω_new, and Ω_all) and benchmarks to quantify knowledge retention and new learning in neural networks.
• It empirically compares five mechanisms, including EWC, PathNet, and GeppNet, across diverse high-dimensional datasets.
• Insights highlight the need for hybrid strategies to bolster incremental learning and prevent knowledge loss in dynamic environments.

Measuring Catastrophic Forgetting in Neural Networks

Catastrophic forgetting remains a significant challenge in the development of neural networks, particularly within the field of incremental learning. The paper "Measuring Catastrophic Forgetting in Neural Networks" by Kemker et al. addresses this issue comprehensively by introducing new metrics, benchmarks, and mechanisms to evaluate and mitigate catastrophic forgetting in MLP-based neural networks.

Overview

The study tackles the limitations of traditional neural networks when tasked with learning sequentially. Once learning a new task commences, neural networks often struggle to retain previously acquired knowledge—a phenomenon termed catastrophic forgetting. The authors categorize methodologies to alleviate this problem into five mechanisms: regularization, ensembling, rehearsal, dual-memory, and sparse-coding.

Key Contributions

1. Empirical Comparisons: The paper undertakes an extensive empirical analysis of the five mechanisms (regularization via EWC, ensembling via PathNet, rehearsal methods with GeppNet, dual-memory using GeppNet+STM, and sparse-coding with FEL) to assess their viability in mitigating forgetting.
2. New Benchmarks: Prior research typically relies on small datasets like MNIST. Here, the authors establish robust benchmarks using varied real-world datasets composed of high-dimensional image/audio data.
3. Novel Metrics: The introduction of specific metrics (Ω_base, Ω_new, and Ω_all) provides a quantifiable method to evaluate both retention of prior knowledge and acquisition of new information.

Experimental Results

Across the board, results indicate significant performance disparities when moving from simple datasets (MNIST) to more complex ones (CUB-200, AudioSet). This underscores the inadequacy of relying solely on small-scale datasets for evaluating catastrophic forgetting solutions. Notably:

• EWC and PathNet demonstrated superior performance, particularly in data permutation scenarios. Their respective mechanisms (weight consolidation and evolutionary path selection) adeptly manage task overlap and weight freezing.
• GeppNet models excelled in incremental class learning, efficiently balancing new knowledge absorption and old knowledge retention through rehearsal strategies.
• Sparse-Coding via FEL showed capacity for preventing forgetting but requires significant computational overhead, presenting challenges for deployment at scale.

Implications

This study is pivotal in setting a foundation for future research focused on overcoming forgetting in neural networks. By presenting concrete benchmarks and metrics, the work presents researchers with tools to rigorously evaluate potential solutions to a longstanding issue in AI. As neural networks increasingly form the backbone of intelligent systems, enhancing their capability to learn incrementally without forgetting is essential for real-world applications, particularly in scenarios involving continual data ingestion and adaptation.

Future Directions

Future research could benefit from hybrid approaches that integrate the strengths of multiple mechanisms presented in this study. Developing efficient algorithms that optimize both memory and computational resources while supporting lifelong learning remains an open challenge. The implication is clear: solving catastrophic forgetting is crucial for deploying adaptive, intelligent agents in dynamic environments.

This paper provides a crucial step towards understanding and mitigating catastrophic forgetting, setting a foundational work for subsequent explorations in incremental learning frameworks.