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..
AnaCP: Toward Upper-Bound Continual Learning via Analytic Contrastive Projection
Authors: Saleh Momeni, Changnan Xiao, Bing Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show that Ana CP not only outperforms existing baselines but also achieves the accuracy level of joint training, which is regarded as the upper bound of CIL. |
| Researcher Affiliation | Academia | Saleh Momeni1, Changnan Xiao2, Bing Liu1 1 Department of Computer Science, University of Illinois Chicago EMAIL EMAIL |
| Pseudocode | No | The paper describes methods and algorithms using mathematical formulas and descriptive text, but it does not contain any explicitly labeled pseudocode or algorithm blocks with structured steps. |
| Open Source Code | Yes | 1The code of Ana CP is available at https://github.com/Saleh Momeni/Ana CP. |
| Open Datasets | Yes | Datasets: We conduct experiments on five publicly available datasets: CIFAR100 (100 classes) [61], Image Net-R (200 classes) [62], CUB (200 classes) [63], Tiny Image Net (200 classes) [64], and Cars (196 classes) [65]. For all datasets, we use the official train and test splits. |
| Dataset Splits | Yes | For all datasets, we use the official train and test splits. Each dataset is split into 10 disjoint tasks by shuffling classes, and experiments are repeated with three different random seeds to account for variability in class-task assignments. |
| Hardware Specification | Yes | All experiments are conducted on a single NVIDIA A100 GPU with 80GB VRAM. |
| Software Dependencies | No | The paper mentions using PTMs like DINO-v2 and Mo Co-v3, but does not specify software dependencies such as programming languages or machine learning frameworks with their version numbers. |
| Experiment Setup | Yes | RP dimension D = 5000, number of CP heads H = 3, and number of generated feature representations per class for the classifier R = 100 as the default configuration; ablations on other variants are included. The regularization parameter λ in Eq. 6 for ELM is set to 102, and the coefficient α for class mean repulsion is set to 1. These values were optimized on a validation set derived from the CIFAR100 training set and are consistently used across all datasets and both PTMs. |