Skip to main content
QUICK REVIEW

[论文解读] Hierarchical Decomposition of Prompt-Based Continual Learning: Rethinking Obscured Sub-optimality

Liyuan Wang, Jingyi Xie|arXiv (Cornell University)|Oct 11, 2023
Domain Adaptation and Few-Shot Learning被引用 18
一句话总结

本论文提出 HiDe-Prompt,一种层次化分解方法,明确在任务内预测、任务身份推断和任务自适应预测方面进行优化,用于基于提示的持续学习,在自监督预训练下达到最先进的结果。

ABSTRACT

Prompt-based continual learning is an emerging direction in leveraging pre-trained knowledge for downstream continual learning, and has almost reached the performance pinnacle under supervised pre-training. However, our empirical research reveals that the current strategies fall short of their full potential under the more realistic self-supervised pre-training, which is essential for handling vast quantities of unlabeled data in practice. This is largely due to the difficulty of task-specific knowledge being incorporated into instructed representations via prompt parameters and predicted by uninstructed representations at test time. To overcome the exposed sub-optimality, we conduct a theoretical analysis of the continual learning objective in the context of pre-training, and decompose it into hierarchical components: within-task prediction, task-identity inference, and task-adaptive prediction. Following these empirical and theoretical insights, we propose Hierarchical Decomposition (HiDe-)Prompt, an innovative approach that explicitly optimizes the hierarchical components with an ensemble of task-specific prompts and statistics of both uninstructed and instructed representations, further with the coordination of a contrastive regularization strategy. Our extensive experiments demonstrate the superior performance of HiDe-Prompt and its robustness to pre-training paradigms in continual learning (e.g., up to 15.01% and 9.61% lead on Split CIFAR-100 and Split ImageNet-R, respectively). Our code is available at \url{https://github.com/thu-ml/HiDe-Prompt}.

研究动机与目标

  • 激励在现实自监督预训练设置下研究基于提示的持续学习。
  • 从理论上将持续学习目标分解为分层组件:任务内预测、任务身份推断、以及任务自适应预测。
  • 提出 HiDe-Prompt,通过任务特异性提示和表示统计显式优化分层组件。
  • 引入对比正则化策略以协调分层组件。
  • 在多个基准上展示经验增益,显示对预训练范式的鲁棒性。

提出的方法

  • 问题被设定为无复现回放的持续学习,包含一个冻结的预训练主干网络和任务特异性提示。
  • 回顾并比较基于提示的方法(ProT 与 PreT),强调如何从无指示的表示中推断任务身份。
  • HiDe-Prompt 扩展一个任务特异性提示池,并使用提示集合将知识迁移到新任务,同时缓解遗忘。
  • 通过专用分支优化 WTP、TII、TAP:WTP 使用带对比正则项的交叉熵,利用旧任务统计;TII 使用辅助的持续自适应输出层从未指示表示中预测任务身份;TAP 使用对所有已见类别适配的输出头。
  • 将每个类别的表征统计建模(以高斯为中心),以实现基于分布的预测;交叉熵损失 H_WTP、H_TII、H_TAP 指导分层优化(方程6–12)。
  • 测试时,模型通过辅助的 TII 路径选择任务身份,然后通过任务特定提示预测标签。

实验结果

研究问题

  • RQ1预训练范式(自监督 vs 监督)如何影响基于提示的持续学习的有效性?
  • RQ2在自监督预训练下,将持续学习目标分解为 WTP、TII、TAP 的分层是否能带来更好表现?
  • RQ3如何组织并规范化任务特异性提示,以在实现知识迁移的同时避免灾难性遗忘?
  • RQ4对未指示/指示表示进行统计建模(如高斯分布)是否能实现跨任务的有效任务身份和类别预测?

主要发现

PTMMethodSplit CIFAR-100 FAASplit CIFAR-100 CAASplit CIFAR-100 FFMSplit ImageNet-R FAASplit ImageNet-R CAASplit ImageNet-R FFM
Sup-21KHiDe-Prompt (Ours)92.61 ± 0.2894.03 ± 0.013.16 ± 0.1075.06 ± 0.1276.60 ± 0.012.17 ± 0.19
Sup-21KL2P [41]83.06 ± 0.1788.25 ± 0.016.58 ± 0.4063.65 ± 0.1267.25 ± 0.027.51 ± 0.17
Sup-21KDualPrompt [40]86.60 ± 0.1990.64 ± 0.014.45 ± 0.1668.79 ± 0.3171.96 ± 0.044.49 ± 0.14
Sup-21KS-Prompt++ [39]88.81 ± 0.1892.25 ± 0.033.87 ± 0.0569.68 ± 0.1272.50 ± 0.043.29 ± 0.05
Sup-21KCODA-Prompt [30] ∗86.94 ± 0.6391.57 ± 0.754.04 ± 0.1870.03 ± 0.4774.26 ± 0.245.17 ± 0.22
iBOT-21KHiDe-Prompt (Ours)93.02 ± 0.1594.56 ± 0.051.33 ± 0.2470.83 ± 0.1773.23 ± 0.082.46 ± 0.21
iBOT-21KL2P [41]79.00 ± 0.2885.13 ± 0.055.55 ± 0.3655.35 ± 0.2858.62 ± 0.053.73 ± 0.53
iBOT-21KDualPrompt [40]78.76 ± 0.2386.16 ± 0.029.84 ± 0.2454.55 ± 0.5358.69 ± 0.015.38 ± 0.70
iBOT-21KS-Prompt++ [39]79.14 ± 0.6585.85 ± 0.179.17 ± 1.3355.16 ± 0.8358.48 ± 0.184.07 ± 0.16
iBOT-21KCODA-Prompt [30]80.83 ± 0.2787.02 ± 0.207.50 ± 0.2561.22 ± 0.3566.76 ± 0.379.66 ± 0.20
iBOT-21KHiDe-Prompt (Ours)93.68 ± 0.1594.56 ± 0.051.21 ± 0.2471.33 ± 0.2173.62 ± 0.132.79 ± 0.26
iBOT-1KHiDe-Prompt (Ours)93.48 ± 0.1195.02 ± 0.011.00 ± 0.2471.33 ± 0.2173.62 ± 0.132.79 ± 0.26
iBOT-1KL2P [41]75.57 ± 0.4182.69 ± 0.067.23 ± 0.9360.97 ± 0.2665.95 ± 0.024.07 ± 0.66
iBOT-1KDualPrompt [40]76.63 ± 0.0585.08 ± 0.128.41 ± 0.4061.51 ± 1.0567.11 ± 0.085.02 ± 0.52
iBOT-1KS-Prompt++ [39]77.53 ± 0.5685.66 ± 0.168.07 ± 0.9760.82 ± 0.6866.03 ± 0.914.16 ± 0.14
iBOT-1KCODA-Prompt [30]79.11 ± 1.0286.21 ± 0.497.69 ± 1.5766.56 ± 0.6873.14 ± 0.577.22 ± 0.38
iBOT-1KHiDe-Prompt (Ours)93.56 ± 0.1294.95 ± 0.041.12 ± 0.2171.21 ± 0.2073.50 ± 0.122.65 ± 0.25
DINO-1KHiDe-Prompt (Ours)92.51 ± 0.1194.25 ± 0.010.99 ± 0.2168.11 ± 0.1871.70 ± 0.013.11 ± 0.17
MoCo-1KHiDe-Prompt (Ours)91.57 ± 0.2093.70 ± 0.011.19 ± 0.1863.77 ± 0.4968.26 ± 0.013.57 ± 0.96
  • 若仅优化提示而不进行分层协调,基于提示的持续学习在自监督预训练下会退化。
  • HiDe-Prompt 在多个基准下达到最先进结果,如 Split CIFAR-100 和 Split ImageNet-R,在不同预训练范式下。
  • 与强基线相比,HiDe-Prompt 显著提升,例如在 Split CIFAR-100 上 FAA 提升最高 15.01%、在 Split ImageNet-R 上提升 9.61%。
  • 使用任务特异性提示集合、提示集合和旧任务统计,以及对比正则化,有助于提升 WTP,并使 TAP 与旧任务对齐。
  • 辅助 TII 和适应的 TAP 头持续改善身份推断和跨任务类别预测,提升 CIL 性能。
  • 在诸如 Sup-21K、iBOT-21K、iBOT-1K、DINO-1K、MoCo-1K 等 PTM 下,HiDe-Prompt 始终优于 L2P、DualPrompt、S-Prompt++、CODA-Prompt(见表1)。

更好的研究,从现在开始

从论文设计到论文写作,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。