Dynamic Feature Learning and Matching for Class-Incremental Learning

TL;DR

The Dynamic Feature Learning and Matching (DFLM) model is proposed, which introduces class weight information and non-stationary functions to extend the mix DA method for dynamically adjusting the focus on memory during training and proposes the matching loss to facilitate the alignment between the learned dynamic features and the classifier by minimizing the distribution distance.

Abstract

Class-incremental learning (CIL) has emerged as a means to learn new classes incrementally without catastrophic forgetting of previous classes. Recently, CIL has undergone a paradigm shift towards dynamic architectures due to their superior performance. However, these models are still limited by the following aspects: (i) Data augmentation (DA), which are tightly coupled with CIL, remains under-explored in dynamic architecture scenarios. (ii) Feature representation. The discriminativeness of dynamic feature are sub-optimal and possess potential for refinement. (iii) Classifier. The misalignment between dynamic feature and classifier constrains the capabilities of the model. To tackle the aforementioned drawbacks, we propose the Dynamic Feature Learning and Matching (DFLM) model in this paper from above three perspectives. Specifically, we firstly introduce class weight information and non-stationary functions to extend the mix DA method for dynamically adjusting the focus on memory during training. Then, von Mises-Fisher (vMF) classifier is employed to effectively model the dynamic feature distribution and implicitly learn their discriminative properties. Finally, the matching loss is proposed to facilitate the alignment between the learned dynamic features and the classifier by minimizing the distribution distance. Extensive experiments on CIL benchmarks validate that our proposed model achieves significant performance improvements over existing methods.

Figures (10)

• Figure 1: The overview of the proposed model.
• Figure 2: (a) The visualization results of sigmoid function $\sigma_{\gamma,\tau}(e)$ with different $\gamma$ and $\tau$ in Eq. \ref{['equation_lambda_hat_rewrite']}. (b)(c)(d) The visualization results of mean functions $\mu(e)$ and $\hat{\mu}(e)$ with intervals and random examples.
• Figure 3: Performance comparison for each step. From left to right: CIFAR-100 B0 10 steps, CIFAR-100 B0 20 steps, CIFAR-100 B50 5 steps, and CIFAR-100 B50 10 steps.
• Figure 4: Performance comparison for each step. From left to right: ImageNet-100 B0 10 steps, ImageNet-100 B0 20 steps, ImageNet-100 B50 5 steps, and ImageNet-100 B50 10 steps.
• Figure 5: Ablation comparison of hyper-parameters $\gamma$ and $\tau$ in MC-Mix on CIFAR-100 benchmark.

Theorems & Definitions (1)

• proof