On-device continual learning (CL) enables lifelong and privacy-preserving learning for various edge intelligent applications. Increasing the number of model parameters as new learning tasks emerge is effective in ensuring learning quality but inefficient in memory cost, especially for resource-constrained devices. In this paper, we introduce InfiniCL, the first ondevice CL system that dynamically balances memory cost and learning quality. A key idea behind InfiniCL is elastic continual learning: selectively freezing layers in the expanding model and periodically distilling the model, preventing unbounded memory growth while preserving learning quality for new tasks. This novel CL paradigm opens a new challenging problem: how to decide the memory allocation of the model and data to achieve better learning Quality of Service (QoS) under the limited memory budget? To alleviate this challenge, we further propose a Bayesian Optimization-driven algorithm to jointly optimize layer freezing selection and data-model memory allocation. Evaluations show that InfiniCL outperforms state-of-the-art methods on diverse memory constraints, achieving 5.34-7.15% and 2.72-9.36% higher accuracy on CIFAR-100 and ImageNet-100, respectively.

Multi-grained cross-modal image-text retrieval models have demonstrated promising outcomes through the alignment of local and global features. However, this advancement often results in larger model sizes and higher computational requirements, which raises concerns regarding the balance between performance and efficiency. To address this challenge, we introduce a novel lightweight multi-grained (LMG) image-text retrieval paradigm aimed at enhancing model efficiency. Specifically, in our approach, we first re-frame the retrieval problem as a cascaded representation learning task. This involves leveraging only fine-grained features to capture coarse-grained constraints, thereby reducing computational burden while maintaining accuracy. Furthermore, we replace computationally expensive parametric feature aggregation methods with three efficient parameter-free alternatives: auto-correlation matrix, discrete linear convolution, and discrete Fourier transform. The proposed LMG model is extensively compared with state-of-the-art approaches on two benchmark datasets, i.e., Flickr30K and MSCOCO, and the experimental results highlight the superior performance of LMG. Additionally, we explore the impact of different feature aggregation methods on LMG and conduct a sensitivity analysis on the coarse and fine-grained constraints ratio hyper-parameter.