Enabling On-Device Learning via Experience Replay with Efficient Dataset Condensation

TL;DR

This work tackles on-device continual learning from unlabeled, non-i.i.d. data streams by condensing incoming information into a compact, informative buffer. It introduces DECO, a framework combining majority-vote pseudo-labeling, efficient one-step gradient matching, and contrastive learning to update a synthetic buffer without discarding existing data. The approach achieves substantial accuracy gains, especially under tight buffer constraints (e.g., IpC=1), while significantly reducing condensation time through outer-loop optimization and finite-difference gradient approximations. The results demonstrate improved robustness, faster learning, and practical scalability for edge devices facing streaming data with limited resources.

Abstract

Upon deployment to edge devices, it is often desirable for a model to further learn from streaming data to improve accuracy. However, extracting representative features from such data is challenging because it is typically unlabeled, non-independent and identically distributed (non-i.i.d), and is seen only once. To mitigate this issue, a common strategy is to maintain a small data buffer on the edge device to hold the most representative data for further learning. As most data is either never stored or quickly discarded, identifying the most representative data to avoid significant information loss becomes critical. In this paper, we propose an on-device framework that addresses this issue by condensing incoming data into more informative samples. Specifically, to effectively handle unlabeled incoming data, we propose a pseudo-labeling technique designed for unlabeled on-device learning environments. Additionally, we develop a dataset condensation technique that only requires little computation resources. To counteract the effects of noisy labels during the condensation process, we further utilize a contrastive learning objective to improve the purity of class data within the buffer. Our empirical results indicate substantial improvements over existing methods, particularly when buffer capacity is severely restricted. For instance, with a buffer capacity of just one sample per class, our method achieves an accuracy that outperforms the best existing baseline by 58.4% on the CIFAR-10 dataset.

Figures (4)

• Figure 1: Overview of DECO. The process begins by labeling and filtering the incoming unlabeled data stream through (a) majority voting. Subsequently, the limited-size synthetic data buffer is updated through (b) efficient on-device condensation and (c) contrastive learning among the synthetic data within the buffer.
• Figure 2: Top 3 classes most frequently misclassified in CIFAR-10 for selected classes.
• Figure 3: Learning curves on different datasets depict the average accuracy in relation to the amount of input data.
• Figure 4: Parameter analysis. (a) shows the effect of varying training intervals $\beta$ on model performance. (b) shows the loss curve for different condensation optimization settings. (c) shows the influence of different filter thresholds $M$ on pseudo-labeling accuracy and classification performance. (d) shows the impact of the loss weighting factor $\alpha$ on average end accuracy.