iTAML: An Incremental Task-Agnostic Meta-learning Approach

Jathushan Rajasegaran; Salman Khan; Munawar Hayat; Fahad Shahbaz Khan; Mubarak Shah

Source-linked AI summary

iTAML: An Incremental Task-Agnostic Meta-learning Approach

Jathushan Rajasegaran, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Mubarak Shah

arXiv:2003.11652v1Mar 25, 2020cs.LGcs.CVstat.ML

TL;DR

Incremental learning requires models to acquire new tasks without losing prior knowledge. iTAML learns shared meta-parameters, balances old and new information, and automatically predicts and adapts to tasks; experiments report consistent improvements across five classification datasets.

Problem
Class-incremental learning must preserve old knowledge while sequentially learning new class groups, but plasticity, stability, catastrophic forgetting, interference, task information, and replay imbalance remain challenges.
Source: §1 · p. 1
Method
iTAML learns task-agnostic shared parameters with an equilibrium-preserving meta-update, automatically predicts the task, and quickly adapts using task-specific parameters.
Source: §1 · p. 2, §3 · p. 3, §3.1 · p. 4
Results
iTAML consistently improves performance across ImageNet, CIFAR100, MNIST, SVHN, and MS-Celeb, including a 21.3% gain over RPS-net on CIFAR100 with 10 incremental tasks.
Source: §4.2 · p. 7, §5 · p. 8
Takeaways & Limitations
The approach provides a generic model that can be rapidly updated for specific tasks while maintaining an equilibrium between old and new task information.
Source: §5 · p. 8
The analysis assumes incorrect predictions are uniformly distributed when deriving the probability of predicting the correct task.
Source: §3.4 · p. 5

Abstract

from arXiv · show

Humans can continuously learn new knowledge as their experience grows. In contrast, previous learning in deep neural networks can quickly fade out when they are trained on a new task. In this paper, we hypothesize this problem can be avoided by learning a set of generalized parameters, that are neither specific to old nor new tasks. In this pursuit, we introduce a novel meta-learning approach that seeks to maintain an equilibrium between all the encountered tasks. This is ensured by a new meta-update rule which avoids catastrophic forgetting. In comparison to previous meta-learning techniques, our approach is task-agnostic. When presented with a continuum of data, our model automatically identifies the task and quickly adapts to it with just a single update. We perform extensive experiments on five datasets in a class-incremental setting, leading to significant improvements over the state of the art methods (e.g., a 21.3% boost on CIFAR100 with 10 incremental tasks). Specifically, on large-scale datasets that generally prove difficult cases for incremental learning, our approach delivers absolute gains as high as 19.1% and 7.4% on ImageNet and MS-Celeb datasets, respectively.

1. Introduction

Class-incremental learning must balance retaining old knowledge with learning new class groups under limited access to prior data. iTAML addresses this with task-agnostic meta-learning that learns shared parameters, predicts tasks automatically, and adapts while reducing forgetting and imbalance.

Incremental learning seeks one model that continually learns new tasks while preserving old knowledge with limited or no access to previous data.Source: §1 · p. 1
Class-incremental learning exposes groups of classes sequentially, creating a tension between plasticity for new tasks and stability for old ones.Source: §1 · p. 1
iTAML learns a generic meta-model whose shared parameters balance current and old knowledge instead of fitting a single one-size-fits-all model.Source: §1 · pp. 1–2
Meta-learning methods such as MAML, FOMAML, and Reptile lack sufficient task adaptation or old-task retention, while task information and replay buffers impose additional constraints.Source: §1 · pp. 1–2
iTAML automatically predicts the task, adapts to it, separates generic features from task-specific classifiers, and tunes task-specific parameters separately to reduce imbalance.Source: Figure 1 · p. 1, §1 · p. 2

2. Related Work

Prior incremental-learning research includes architectural, rehearsal, regularization, and meta-learning strategies. Meta-learning methods aim to learn rapid adaptation, but existing approaches impose restrictions involving gradient alignment, task streams, inference data, or single-sample updates.

Incremental-learning methods modify architectures, rehearse exemplars or generated samples, or operate algorithmically through regularization and meta-learning.Source: §2 · p. 2
Regularization approaches reduce forgetting through distillation, attention preservation, classifier bias correction, synaptic-importance estimates, parameter merging, or Fisher-based constraints.Source: §2 · p. 2
Meta-learning aims to train models across tasks so new tasks can be learned quickly with minimal supervision.Source: §2 · p. 2
Prior meta-learning approaches align updates or disentangle representations and task heads, but rely on fixed objectives, correlated streams, concurrent inference samples, or single-sample inner-loop updates.Source: §2 · p. 2

3. Proposed Method

iTAML learns generic parameters that adapt across sequential tasks while separating shared features from task-specific classifiers. Its meta-update combines task-specific models with task-dependent momentum, enabling task identification and rapid class prediction without inference-time gradient updates.

Core design

iTAML learns shared feature parameters and task-specific classifiers so one generic model can adapt across tasks without requiring task information.The approach is model- and task-agnostic and seeks parameters that work well for new tasks with minor local changes.Source: §3 · p. 3, §3.1 · p. 3

Inner loop

Training samples from current data and exemplar memory are grouped by task into micro-batches, each producing a task-specific model for the outer update.The feature parameters are updated for every task, while each task classifier is updated only on its corresponding micro-batch.Source: §3.1 · p. 3, Figure 2 · p. 3

Outer loop

The outer loop moves the generic model toward the average direction of task-specific updates using a dynamic controller η.This combines the task-specific models rather than optimizing a single task-specific solution.Source: §3.1 · p. 3, Figure 3 · p. 4

Momentum control

η = exp(−βt/T) slows parameter movement as more tasks are observed, balancing adaptation to new tasks with preservation of previous information.The controller is set using a validation-set decay rate β; later-task updates therefore move more slowly than early-task updates.Source: §3.1 · pp. 3–4

Meta-update

Unlike Reptile, iTAML can work with r ≥ 1 inner-loop updates and uses one micro-batch per task to estimate the outer-loop update.The authors motivate this choice by the under-representation of old classes in exemplar-based incremental learning.Source: §3.2 · p. 4

Inference

At inference, iTAML predicts the task from a data continuum, then adapts generic parameters with a small exemplar set to predict classes.Task prediction can reach up to 100% accuracy, while class prediction can be close to joint training without gradient updates after meta-training.Source: §3.2 · pp. 4–5

4. Experiments and Results

Experiments evaluate iTAML across incremental-learning benchmarks, meta-learning settings, and ablations. The method consistently outperforms competing approaches, including large gains on CIFAR100 and large-scale datasets.

Benchmark setup

iTAML is evaluated on MNIST, SVHN, CIFAR100, ImageNet-100, ImageNet-1K, and MS-Celeb-10K incremental-learning benchmarks.The experiments use varied dataset scales, task counts, architectures, and exemplar-memory sizes.Source: §4.1 · p. 6

Meta-learning comparisons

77.79% average classification accuracy is maintained after 10 CIFAR100 tasks, while task prediction remains above 95% in the task-agnostic setting.The comparison includes FOMAML, Reptile, joint training, and iTAML under task-agnostic, task-aware, and no-inference-update conditions.Source: §4.2 · p. 6, Figure 5 · p. 7

CIFAR100 results

21.3% is iTAML’s margin over RPS-net when incrementally learning 10 CIFAR100 tasks; gains are 23.6% and 18.2% for 5 and 20 classes, respectively.The experiments report state-of-the-art performance across all three CIFAR100 incremental settings.Source: §4.2 · p. 7, Figure 6 · p. 8

Large-scale results

19.1% is iTAML’s absolute gain on ImageNet-1K, while it reaches 95.02% accuracy on MS-Celeb-10K with 10,000 classes.It also achieves 89.8% on ImageNet-100, surpassing the current best method by 15.7%.Source: §4.2 · p. 7, Table 2 · p. 8

Ablation analysis

Increasing the data-continuum size improves task-prediction accuracy by attenuating noise in the average response.The ablations examine continuum size, inner-loop updates, β, and dataset-specific task and class accuracy.Source: §4.3 · p. 7, Figure 7 · p. 8

5. Conclusion

iTAML trains a generic model that adapts quickly to specific tasks while preserving an equilibrium between old and new task information. Its task-agnostic design automatically detects the task and improves classification consistently across several datasets.

iTAML optimizes common parameters across encountered tasks rather than specializing the model to a single task.Source: §5 · p. 8
The meta-learning approach rapidly updates a generic model for a specific task while balancing old and new task information.Source: §5 · p. 8
iTAML automatically detects the task at inference and updates itself to perform well on the given inputs.Source: §5 · p. 8
Experiments show consistent improvements across ImageNet, CIFAR100, MNIST, SVHN and MS-Celeb.Source: §5 · p. 8

Appendix A.1. iTAML vs Other Meta Algorithms

The appendix formalizes iTAML's inner-loop updates and meta-gradient construction, including a formulation that permits any number of inner-loop updates. It contrasts this with Reptile's requirement for more than one update.

The inner-loop parameters follow gradient descent on a task loss using learning rate α.Source: Appendix A.1 · p. 11
The meta-gradient is defined from the displacement between initial and updated parameters, following a Reptile-style update rule.Source: Appendix A.1 · p. 11
iTAML permits r ≥ 1 inner-loop updates, whereas Reptile requires r > 1 because r = 1 becomes joint training.Source: Appendix A.1 · p. 11
For task t, the meta-update uses the gradients available from the task sequence.Source: Appendix A.1 · p. 11
iTAML updates shared feature parameters and task-specific classification parameters during each task's inner loop.Source: Appendix A.1 · p. 11

Appendix A.2. Additional Results

Additional experiments examine adaptation settings, dataset-specific choices, model size, and optimizer modularity. The results show limited benefit from increasing adaptation epochs and strong performance across small and full backends and multiple optimizers.

Figure 8 evaluates CIFAR100 classification accuracy with 2 tasks, 2000 exemplar samples, ResNet-18(1/3), and p=20.Source: Figure 8 · p. 11
b=1 is used for adaptation because increasing b to 5 does not significantly improve accuracy while increasing time complexity linearly.Source: Appendix A.2 · p. 11
For SVHN, r=4 is retained for the last task because lower data-distribution variance can trap the model at early local minima.Source: Appendix A.2 · p. 11
69.94% accuracy is achieved with a 0.49M-parameter model, a 13.46% gain over the 77.5M-parameter RPS-net method.Source: Appendix A.2 · p. 12
ResNet-18 reaches 80.27% accuracy, while SGD, Adam and RAdam produce 70.34%, 74.83% and 76.63% accuracy, respectively.Source: Appendix A.2 · p. 12