Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Gradient-Guided Epsilon Constraint Method for Online Continual Learning
Authors: Song Lai, Changyi Ma, Fei Zhu, Zhe Zhao, Xi Lin, GAOFENG MENG, Qingfu Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical evaluations on standard OCL benchmarks demonstrate GEC s ability to achieve a superior trade-off, leading to improved overall performance. Code is available at https://github.com/laisong-22004009/GEC_OCL. Our empirical analysis (Section 4) further explores this relationship. We find that ER s approach, which allows for some beneficial conflicts with memory task gradients, can lead to better generalization than strictly adhering to memory constraints. 6 Experiments 6.1 Experiment Setup Evaluation Metrics. 6.2 Experimental Results |
| Researcher Affiliation | Academia | 1Department of Computer Science, City University of Hong Kong 2Centre for Artificial Intelligence and Robotics, HKISI, CAS 3Institute of Automation, Chinese Academy of Sciences 4School of Artificial Intelligence, University of Chinese Academy of Sciences EMAIL, EMAIL |
| Pseudocode | Yes | D Algorithm Details This section provides the pseudo-code for the proposed GEC method and a list of key symbols used. Algorithm 1 Gradient-Guided Epsilon Constraint (GEC) |
| Open Source Code | Yes | Empirical evaluations on standard OCL benchmarks demonstrate GEC s ability to achieve a superior trade-off, leading to improved overall performance. Code is available at https://github.com/laisong-22004009/GEC_OCL. |
| Open Datasets | Yes | Following [Buzzega et al., 2020, Arani et al., 2022, Lai et al., 2025], we evaluate our method on three commonly used benchmarks for online continual learning: Sequential CIFAR-10 (Seq-CIFAR10), Sequential CIFAR-100 (Seq-CIFAR100) [Krizhevsky et al., 2009], and Sequential Tiny Image Net (Seq-Tiny Image Net) [Buzzega et al., 2020]. |
| Dataset Splits | Yes | For Seq-CIFAR10, the dataset is split into 5 tasks, each containing 2 classes. We test with memory buffer sizes (|M|) of 0.6k and 1k samples. For Seq-CIFAR100, the dataset is split into 20 tasks, each with 5 classes, using memory buffer sizes of 1k and 5k samples. Seq-Tiny Image Net is split into 20 tasks (10 classes per task), with memory buffer sizes of 2k and 5k. Table 9: Details of datasets Dataset Total Classes Tasks (N) Classes per Task Training Images Test Images Resolution Seq-CIFAR10 10 5 2 50,000 10,000 32 32 3 Seq-CIFAR100 100 20 5 50,000 10,000 32 32 3 Seq-Tiny Image Net 200 20 10 100,000 10,000 32 32 3 |
| Hardware Specification | Yes | The experiments were performed on 8 NVIDIA A100 GPUs. |
| Software Dependencies | No | The paper mentions "SGD" as the optimizer and "Reduced Res Net-18" as the backbone network, but does not provide specific version numbers for software libraries, frameworks, or programming languages (e.g., Python, PyTorch). |
| Experiment Setup | Yes | The models are trained using Stochastic Gradient Descent (SGD) with a batch size of 32. All experiments are conducted in the online class-incremental learning setting, where each task is trained for a single epoch, and task identity is not available during inference. All reported results are averaged over 5 independent runs. Implementation Details. Consistent with [Chrysakis and Moens, 2023], we use the Reduced Res Net-18 [He et al., 2016] as the backbone network for all experiments. Models are trained using Stochastic Gradient Descent (SGD) with a batch size of 32. The learning rate is set to 0.03 for GEC and all baseline methods. The streaming batch size and replay batch size are set to 32 for all experiments. The memory buffer M is updated using reservoir sampling. GEC Hyperparameters. For our GEC method, the barrier strength parameter α is set to 0.5, and the stability constant δ is set to 10−6. |