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..
Contrastive Consolidation of Top-Down Modulations Achieves Sparsely Supervised Continual Learning
Authors: Viet Anh Khoa Tran, Emre Neftci, Willem Wybo
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show improvements in both class-incremental and transfer learning over state-of-the-art unsupervised approaches, as well as over comparable supervised approaches, using as few as 1% of available labels. |
| Researcher Affiliation | Academia | Viet Anh Khoa Tran Peter Grünberg Institute Forschungszentrum Jülich & RWTH Aachen EMAIL Emre Neftci Peter Grünberg Institute Forschungszentrum Jülich & RWTH Aachen EMAIL Willem A. M. Wybo Peter Grünberg Institute Forschungszentrum Jülich EMAIL |
| Pseudocode | Yes | C Python-style Pseudocode for TMCL |
| Open Source Code | Yes | The code for our experiments is available at https://github.com/tran-khoa/tmcl. |
| Open Datasets | Yes | Our analysis focuses on the CIFAR-100 dataset [117] and the Image Net-100 dataset [111]. For the transfer learning experiments, we perform k NN evaluation on Aircraft [118], CIFAR-10 [117], CUBirds [122], DTD [109], Euro SAT [113], GTSRB [114], STL-10 [110], SVHN [119], and VGGFlower [120]. |
| Dataset Splits | Yes | We adopt a standard class-incremental continual learning protocol on both CIFAR-100 and Image Net-100, dividing the dataset into five sessions, each containing 20 disjoint classes. ... For all CIFAR-100 experiments, we use the same class split as in Ca SSLe, i.e. the same across all seeds. For the Image Net-100 as well as the 10 session experiments (Section F), we use different class splits per seed. |
| Hardware Specification | Yes | We run our experiments on the JUWELS-Booster [115] and JURECA [116] clusters at Forschungszentrum Jülich. For both systems, we use a single NVIDIA A100 GPU per experiment. |
| Software Dependencies | No | We implement our methods based on Py Torch [108] with the Lightning framework [112]. For the augmentations on CIFAR-100, we used kornia [121]. Backbone implementations are adapted from the timm library [123]. |
| Experiment Setup | Yes | For Con Vi T experiments, we use the Adam W optimizer [106] with a batch size of 256 and feedforward weight decay of 0.0001 for all experiments, using a per-session cosine learning rate decay with 10 warmup epochs. Con Vi T experiments are trained with a feedforward learning rate of 0.001 and a modulation learning rate of 0.01. The Res Net experiments are trained with a feedforward learning rate of 1.0 and a modulation learning rate of 0.3 using the LARS optimizer (η = 0.02). The Barlow Twins losses are scaled down by a factor of 0.1 for Con Vi T experiments, and by 0.025 for Res Net experiments. We pick the redundancy-reduction weighting factor λBT = 0.005 for all experiments. |