[Paper Review] Single-Net Continual Learning with Progressive Segmented Training (PST)

Xiaocong Du, Gouranga Charan+2 more|arXiv (Cornell University)|May 28, 2019

Domain Adaptation and Few-Shot Learning39 references4 citations

TL;DR

This paper proposes Progressive Segmented Training (PST), a single-network continual learning method that prevents catastrophic forgetting by segmenting model parameters into important (frozen) and secondary (retained) groups, while using a fixed-size memory buffer for rehearsal. PST achieves state-of-the-art single-head accuracy on CIFAR-10 and CIFAR-100 with significantly reduced FLOPs—over 24× lower in weight update paths—making it highly efficient for edge deployment.

ABSTRACT

There is an increasing need of continual learning in dynamic systems, such as the self-driving vehicle, the surveillance drone, and the robotic system. Such a system requires learning from the data stream, training the model to preserve previous information and adapt to a new task, and generating a single-headed vector for future inference. Different from previous approaches with dynamic structures, this work focuses on a single network and model segmentation to prevent catastrophic forgetting. Leveraging the redundant capacity of a single network, model parameters for each task are separated into two groups: one important group which is frozen to preserve current knowledge, and secondary group to be saved (not pruned) for a future learning. A fixed-size memory containing a small amount of previously seen data is further adopted to assist the training. Without additional regularization, the simple yet effective approach of PST successfully incorporates multiple tasks and achieves the state-of-the-art accuracy in the single-head evaluation on CIFAR-10 and CIFAR-100 datasets. Moreover, the segmented training significantly improves computation efficiency in continual learning.

Motivation & Objective

• Address catastrophic forgetting in continual learning for dynamic edge systems like self-driving vehicles and drones.
• Maintain high accuracy in single-head evaluation without prior task identification, a key challenge in real-world deployment.
• Reduce computational cost in continual learning to enable efficient on-device adaptation with limited resources.
• Preserve prior knowledge using a single static network architecture, avoiding dynamic structure expansion.
• Achieve high efficiency and performance through parameter segmentation and memory-assisted training without additional regularization.

Proposed method

• Segment model parameters into two groups: an important group that is frozen to preserve knowledge and a secondary group that is retained for future learning.
• Use importance sampling to identify and freeze the most critical parameters for prior tasks.
• Maintain a fixed-size memory buffer containing a small subset of previously seen data for continual rehearsal during training.
• Apply memory-assisted balancing to ensure stable learning by dynamically adjusting the contribution of memory samples.
• Progressively train on new tasks using the segmented network and memory buffer, avoiding parameter interference.
• Leverage the redundant capacity of a single network to avoid structural changes while maintaining performance.

Experimental results

Research questions

• RQ1 Can a single, static neural network architecture achieve state-of-the-art performance in continual learning without dynamic architecture expansion?
• RQ2 How effective is parameter segmentation combined with memory rehearsal in mitigating catastrophic forgetting in single-head evaluation?
• RQ3 To what extent does progressive segmentation reduce computational cost in continual learning on edge devices?
• RQ4 What is the relative contribution of importance sampling, model segmentation, and memory balancing to overall performance?
• RQ5 Can PST maintain high accuracy with minimal memory usage and low FLOP overhead in resource-constrained environments?

Key findings

• PST achieves state-of-the-art single-head accuracy on CIFAR-10 and CIFAR-100 across 5, 10, and 20 incremental tasks.
• Removing importance sampling or model segmentation causes a significant accuracy drop of 0.32–0.45 percentage points, demonstrating their critical role.
• Memory-assisted balancing contributes less but still improves performance, with a 0.06–0.11% accuracy gain under varying memory budgets.
• PST reduces FLOPs in the weight update path by over 24× compared to regularization methods like iCaRL, especially beneficial for edge deployment.
• The method maintains high accuracy even with limited memory, though performance gains saturate beyond a certain memory size.
• The segmented training approach enables faster training with lower latency, particularly due to reduced operations in the weight update path.