Lifelong Compression Mixture Model via Knowledge Relationship Graph
Authors: Fei Ye, Adrian G. Bors
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on different datasets show that LGMM achieves the best performance for TFCL. |
| 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 steps of its method in text and with diagrams (Figure 1, 2) but does not provide a formal pseudocode or algorithm block. |
| Open Source Code | Yes | Supplementary materials (SM) and source code are available1. 1https://github.com/dtuzi123/Lifelong Compression Mix |
| Open Datasets | Yes | Datasets : Split MNIST divides MNIST (Le Cun et al. 1998) containing 60k training samples, into five tasks according to pairs of digits in increasing order (De Lange and Tuytelaars 2021). Split CIFAR10 splits CIFAR10 (Krizhevsky and Hinton 2009) into five tasks where each task consists images from two different classes (De Lange and Tuytelaars 2021). Split CIFAR100 divides CIFAR100 into 20 tasks where each task has 2500 samples from 5 different classes (Lopez-Paz and Ranzato 2017). |
| Dataset Splits | No | The paper mentions using "training samples" and evaluating against a "test dataset", but does not explicitly specify distinct validation dataset splits with percentages or sample counts for reproducibility. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware used to run its experiments (e.g., GPU/CPU models, memory amounts). |
| Software Dependencies | No | The paper does not provide specific version numbers for key software components or libraries needed to reproduce the experiments. |
| Experiment Setup | Yes | Network architecture and hyperparameters : In the experiments we adapt the setting from (De Lange and Tuytelaars 2021), where Res Net-18 (He et al. 2016) is used as the classifier for Split CIFAR10 and Split CIFAR100. We also use an MLP network with 2 hidden layers of 400 units (De Lange and Tuytelaars 2021) as the classifier for Split MNIST. The maximum memory size for Split MNIST, Split CIFAR10, and Split CIFAR100 is 2000, 1000 and 5000, respectively. In each training step Ti, we only access a small batch (b = 10) of training samples. The number of epochs for each training step/time is 10. The model expansion threshold λ from Eq. (7) for Split MNIST, Split CIFAR10, Split CIFAR100, and Split MImage Net is 0.009, 0.04, 0.03 and 0.055, respectively. |