Privacy-Aware Lifelong Learning

Authors: Ozan Özdenizci, Elmar Rueckert, Robert Legenstein

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically demonstrate the scalability of PALL across various architectures in image classification, and provide a state-of-the-art solution that uniquely integrates lifelong learning and privacyaware unlearning mechanisms for responsible AI applications. 5 EXPERIMENTAL RESULTS In Table 1 we evaluate our approach against state-of-the-art methods in lifelong learning, by extending various methods to the experimental setting of task incremental learning and unlearning.
Researcher Affiliation Academia Ozan Ozdenizci1, Elmar Rueckert1, Robert Legenstein2 1 Chair of Cyber-Physical-Systems, Montanuniversit at Leoben, Austria 2 Institute of Machine Learning and Neural Computation, Graz University of Technology, Austria EMAIL EMAIL
Pseudocode Yes Our proposed methodology for privacy-aware lifelong learning is outlined in Algorithm 1. Algorithm 1 Privacy-Aware Lifelong Learning (PALL)
Open Source Code Yes Our code is available at: https://github.com/oozdenizci/PALL. Our algorithm is also outlined in Appendix A.2, and our implementations are available at: https://github.com/oozdenizci/PALL.
Open Datasets Yes We performed experiments with sequential CIFAR10 (S-CIFAR10: 5 tasks 2 classes), CIFAR100 (S-CIFAR100: 10 tasks 10 classes) and Tiny Image Net (S-Tiny Image Net: 20 tasks 10 classes, 40 tasks 5 classes, or 100 tasks 2 classes) datasets. We used Res Net-18 and Res Net-34 models in S-CIFAR10/100 experiments which are commonly used as benchmarks in lifelong learning, and attention-based Vi T-T/8 architectures with S-Tiny Image Net (see Appendix A.1 for details).
Dataset Splits Yes We experimented with S-CIFAR10: 5 tasks 2 classes, S-CIFAR100: 10 tasks 10 classes (Krizhevsky, 2009), and S-Tiny Image Net in various sequential configurations: 20 tasks 10 classes, 40 tasks 5 classes, and 100 tasks 2 classes (Le & Yang, 2015). To adapt these datasets for our experimental setting, we first randomly allocate the classes in the dataset into disjoint tasks, and subsequently generate user request sequences R1:r consisting of T task learning and Nu task unlearning instructions arranged in a logically consistent manner. This process was repeated in a randomized manner for each random seed (e.g., seed 0 to 19).
Hardware Specification Yes We compare training times of each algorithm in Table A7 using an NVIDIA A40 GPU. We evaluate all algorithms on the same test sequence that begins as: R1:r = [(1, D1, L), (2, D2, L), (1, U), . . .].
Software Dependencies No The paper mentions optimizers like SGD and Adam, but does not provide specific version numbers for software libraries or frameworks (e.g., PyTorch, TensorFlow, Python version).
Experiment Setup Yes Training Configurations: We use a stochastic gradient descent (SGD) optimizer with momentum for 20 epochs per S-CIFAR10/100 task learning instruction, with a batch size of 32, learning rate of 0.01, and weight decay with parameter 0.0005. For S-Tiny Image Net, we use an Adam optimizer for 100 epochs with a batch size of 256, and a cosine annealing learning rate scheduler with an initial value of 0.001. Here, we do not use weight decay but instead apply dropout to intermediate activations of Vi T-T/8 with p = 0.1 (Steiner et al., 2022). All methods requiring a memory buffer had a total capacity of 500 and 1000 samples (evenly split across tasks) in S-CIFAR and S-Tiny Image Net experiments, respectively. Unless otherwise specified, for Eq. (6) we use Nf = 50 and β = 0.5 (see Appendix A.2 for further details).