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..
Model Inversion with Layer-Specific Modeling and Alignment for Data-Free Continual Learning
Authors: Ruilin Tong, Haodong Lu, Yuhang Liu, Dong Gong
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on both Res Netand CLIP-based CL settings under the challenging classincremental setting, where the model must classify all classes without access to task identity at inference. In all experiments presented in this section, synthetic samples are generated using our PMI+full-model inversion strategy. Detailed experimental settings are provided in Appendix I. Model inversion for Res Net-based CL. To demonstrate the effectiveness and efficiency of our proposed method, we conduct continual learning experiments on CIFAR-100 [23] and Tiny-Image Net [57], following the experimental setup of R-DFCIL [11] using a Res Net-32 [12] backbone. The classes from each dataset are evenly divided into 5, 10, and 20 disjoint tasks. We compare our method against competitive data-free baselines, including Deep Inversion [60], ABD [39], and R-DFCIL [11]. ... Final average accuracy after training on all tasks is used as the evaluation metric, and results are averaged over multiple runs with different random seeds, as shown in Table 1. |
| Researcher Affiliation | Academia | 1 School of Computer Science and Engineering, University of New South Wales 2 Australian Institute for Machine Learning, The University of Adelaide EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1: PMI+full-model inversion Input: Trained model with parameter θ, Class feature for inversion o L. Output: Synthetic input ˆx for l = L to 1 do ... Algorithm 2: Contrastive feature selection. Input: Gaussian distribution N(µc, σc) of class c, contrastive model fcont(ϕ), selection steps n Output: Feature set for inversion Sfeat Initialize Sfeat from N(µc, σc); for i = 1 to n do ... |
| Open Source Code | Yes | Code is available at https://github.com/Ruilin Tong/PMI-CFS-DFCL. |
| Open Datasets | Yes | Our experiments include the CIFAR-100, Tiny-Image Net, Image Net-R, and CUB-200 datasets. The URLs for these datasets are: https://www.cs.toronto.edu/~kriz/cifar.html https://github.com/hendrycks/imagenet-r https://www.vision.caltech.edu/datasets/cub_200_2011/ |
| Dataset Splits | Yes | The CIFAR-100 dataset contains 100 classes, which we evenly divide into 5, 10, and 20 disjoint tasks. Similarly, the Tiny-Image Net dataset, consisting of 200 classes, is split into 5, 10, and 20 disjoint tasks. ... We maintain an incremental buffer containing a t + b synthetic images, where t denotes the task index, a is the number of new samples generated for each task, and b represents the base samples. The configurations for different settings are provided in Table 14. |
| Hardware Specification | Yes | All experiments are conducted on NVIDIA Ge Force RTX 3090 GPUs with 24 GB of memory, using an Intel(R) Core(TM) i9-12900K CPU. ... All experiments are conducted on a system with four NVIDIA Ge Force RTX 3090 GPUs (24 GB each) and an Intel(R) Core(TM) i9-12900K CPU with 64 GB of RAM. |
| Software Dependencies | No | Our Res Net-based CL experiments are implemented based on the open-source code of R-DFCIL [11] and DCMI [33]. ... Our CLIP-based CL experiments are implemented based on the open-source code of PROOF [68], CODA-Prompt [40], and Mo E-Adapter [61]. |
| Experiment Setup | Yes | Each task takes totally 120 epochs for training, and learning rate is set to 0.01 with 0.1 times decreasing after 60 and 90 epochs, we use SGD optimizer for all the experiments. Loss factors for hard knowledge distillation, relational knowledge distillation and classification head finetuning loss are set to 0.15, 0.5 and 1.5 respectively. ... For model inversion, we use the Adam optimizer, with a learning rate of 0.8 for PMI and 0.4 for full-model inversion. The number of update steps is set to 50 for PMI and 160 for full-model inversion. ... Table 15: Training hyper-parameters for CLIP-based CL, where LR denotes learning rate. |