Dynamic Expansion Diffusion Learning for Lifelong Generative Modelling
Authors: Fei Ye, Adrian G. Bors, Kun Zhang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments Results for Class-Incremental Learning We consider the class-incremental learning of the Split MNIST, Split Fashion, Split SVHN and Split CIFAR10, where each learning task consists of data from 2 consecutive classes of the original datasets MNIST, Fashion, SVHN and CIFAR10, as in (Aljundi, Kelchtermans, and Tuytelaars 2019), and the results are given in Tab. 1. For comparison, we also train a VAE as a Student module when considering CNDPM (Lee et al. 2020) as a Teacher and we evaluate image reconstruction and generation. From Tab. 1 DEDM outperforms other methods when evaluating image reconstructions using PSNR and SSIM. |
| Researcher Affiliation | Academia | 1School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu 2Department of Computer Science, University of York, York YO10 5GH, UK 3MBZUAI, Abu Dhabi, UAE, 4Carnegie Mellon University, Pittsburgh, PA, USA EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1: Training algorithm 1: for i < n do 2: First step (Checking the expansion) : 3: if |Mi 1| |Mmax| then 4: if Eq. (8) is satisfied then 5: Build a new component 6: Clear up the memory buffer Mi 1 7: end if 8: end if 9: Second step (Training process) : |
| Open Source Code | Yes | Code https://github.com/dtuzi123/DEDM |
| Open Datasets | Yes | Datasets DEDM Reservoir-VAE Reservoir-DDPM CGKD-GAN CNDPM CGKD-WVAE CGKD-VAE Split MNIST 34.26 56.10 65.06 56.61 65.34 48.57 46.47 Split Fashion 53.78 107.92 81.15 86.51 175.38 88.92 89.76 Split SVHN 54.79 66.24 86.23 103.91 153.36 101.25 104.29 Split CIFAR10 82.12 156.71 107.25 112.29 234.08 163.52 164.74 |
| Dataset Splits | Yes | We consider the class-incremental learning of the Split MNIST, Split Fashion, Split SVHN and Split CIFAR10, where each learning task consists of data from 2 consecutive classes of the original datasets MNIST, Fashion, SVHN and CIFAR10, as in (Aljundi, Kelchtermans, and Tuytelaars 2019) |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details, such as library or solver names with version numbers. |
| Experiment Setup | Yes | In the experiments, we consider λ = 0.001 to ensure that the first term from the IDDPM loss, Eq. (1), does not overwhelm the second term. ... γ [0, 40] is a hyperparameter that controls the expansion process. ... The maximum memory size of various models for all datasets is 2,000. The number of training epochs and the batch size in each training time are 10 and 64, respectively. |