Lifelong Generative Modelling Using Dynamic Expansion Graph Model

Authors: Fei Ye, Adrian G. Bors8857-8865

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental Results Unsupervised Lifelong Learning Benchmark Setting. We define a novel benchmark for the log-likelihood estimation under LLL, explained in Appendix-K from SM1. We consider learning multiple tasks defined within a single domain, such as MNIST (Le Cun et al. 1998) and Fashion (Xiao, Rasul, and Vollgraf 2017). Following from (Burda, Grosse, and Salakhutdinov 2015) we divide MNIST and Fashion into five tasks (Zenke, Poole, and Ganguli 2017), called Split MNIST (S-M) and Split Fashion (S-F). We use the cross-domain setting where we aim to learn a sequence of domains, called COFMI, consisting of Caltech 101 (Fei-Fei, Fergus, and Perona 2007), OMNIGLOT (Lake, Salakhutdinov, and Tenenbaum 2015), Fashion, MNIST, Inverse Fashion (IFashion) where each task is associated with a distinct dataset. All databases images are binarized.
Researcher Affiliation Academia Fei Ye and Adrian G. Bors Department of Computer Science, University of York, York YO10 5GH, UK fy689@york.ac.uk, adrian.bors@york.ac.uk
Pseudocode No The paper describes the model and its training procedure in text and mathematical formulas but does not include any pseudocode or algorithm blocks.
Open Source Code Yes Supplementary materials (SM) and source code are available1. 1https://github.com/dtuzi123/Expansion-Graph-Model
Open Datasets Yes We consider learning multiple tasks defined within a single domain, such as MNIST (Le Cun et al. 1998) and Fashion (Xiao, Rasul, and Vollgraf 2017). ... We use the cross-domain setting where we aim to learn a sequence of domains, called COFMI, consisting of Caltech 101 (Fei-Fei, Fergus, and Perona 2007), OMNIGLOT (Lake, Salakhutdinov, and Tenenbaum 2015), Fashion, MNIST, Inverse Fashion (IFashion)... We train various models under CCCOSCZC lifelong learning setting, where each task is associated with one of the datasets: Celeb A (Liu et al. 2015), CACD (Chen, Chen, and Hsu 2014), 3D-Chair (Aubry et al. 2014), Ommiglot (Lake, Salakhutdinov, and Tenenbaum 2015), Image Net* (Krizhevsky, Sutskever, and Hinton 2012), Car (Yang et al. 2015), Zappos (Yu and Grauman 2017), CUB (Wah et al. 2010).
Dataset Splits No For a given sequence of tasks {T1, . . . , TN} we consider that each Ti is associated with an unlabeled training set QS i and an unlabeled testing set QT i . The model only sees a sequence of training sets {QS 1 , . . . , QS N} while it is evaluated on {QT 1 , . . . , QT N}.
Hardware Specification No The paper does not specify the hardware used for experiments (e.g., GPU models, CPU types).
Software Dependencies No The paper does not provide specific software dependency versions (e.g., Python, PyTorch, or other library versions).
Experiment Setup No The paper mentions a threshold 'τ = 600' for DEGM's component addition. However, it lacks comprehensive details on the experimental setup such as specific hyperparameters (e.g., learning rate, batch size, epochs, optimizer settings) for training the VAE models.