Sequence Transferability and Task Order Selection in Continual Learning

TL;DR

This work tackles how the order and relationships among tasks in continual learning (CL) affect model performance. It introduces two sequence transferability measures, Total Forward Transferability ($\mathrm{tft}$) and Total Reverse Transferability ($\mathrm{trt}$), built on a base transferability metric (e.g., LogME) to quantify forward learning ease and forgetting across a task sequence. Based on these measures, it proposes the Heuristic Continual Task Order Selection (HCTOS) algorithm to select task orders that improve average CL accuracy compared to random ordering, and demonstrates that TFT/TRT correlate with final performance across multiple CL methods and benchmarks. The results show that TFT/TRT are practical proxies for sequence hardness and that HCTOS is robust to sample size and metric choice, with tangible gains in both single-batch and multi-batch task arrival scenarios.

Abstract

In continual learning, understanding the properties of task sequences and their relationships to model performance is important for developing advanced algorithms with better accuracy. However, efforts in this direction remain underdeveloped despite encouraging progress in methodology development. In this work, we investigate the impacts of sequence transferability on continual learning and propose two novel measures that capture the total transferability of a task sequence, either in the forward or backward direction. Based on the empirical properties of these measures, we then develop a new method for the task order selection problem in continual learning. Our method can be shown to offer a better performance than the conventional strategy of random task selection.

Figures (5)

• Figure 1: The AA, TFT, and TRT of different replay-based CL algorithms over three benchmarks. Given a buffer size, TFT and TRT show the same trend as AA.
• Figure 2: Comparison of AA on task sequences selected by HCTOS and random strategies. Our HCTOS method shows better performance in both single-batch and multiple-batch settings.
• Figure 3: Effects of sample size when training simple models on the performance of HCTOS. Our method is robust to the sample size and is consistently better than the random baseline.
• Figure 4: Average accuracy of HCTOS with respect to four different base transferability metrics. LogME exhibits higher performance than its counterparts.
• Figure 5: The AA, TFT, and TRT of different CL algorithms on Split mutual-CIFAR-10 and Split CIFAR-100. Error bars are the standard deviations of the experiments.

Theorems & Definitions (2)

• Definition 1
• Definition 2