Continual learning via probabilistic exchangeable sequence modelling
Authors: Hanwen Xing, Christopher Yau
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that CL-BRUNO outperforms existing methods on both natural image and biomedical data sets, confirming its effectiveness in real-world applications. |
| Researcher Affiliation | Academia | Hanwen Xing EMAIL Nuffield Department of Women s and Reproductive Health University of Oxford Christopher Yau EMAIL Nuffield Department of Women s and Reproductive Health University of Oxford & Health Data Research UK |
| Pseudocode | Yes | Algorithm 1: CL-BRUNO for TIL |
| Open Source Code | Yes | 1Code reproducing the reported results can be found in https://github.com/hwxing2357/cl_bruno. |
| Open Datasets | Yes | For class-IL (CIL), we use three public datasets: incremental CIFAR10 (i CIFAR-10) (Lopez-Paz and Ranzato, 2017), incremental CIFAR100(i CIFAR-100) (Zenke et al., 2017), and sequential Tiny Image Net (S-Tiny Image Net) (Chaudhry et al., 2019). For task-IL(TIL), we use the MNIST dataset (Le Cun, 1998). We here demonstrate CL-BRUNO under a CIL scenario using the Pan-Cancer Atlas (PANCAN) dataset (Hoadley et al., 2018). We also tested CL-BRUNO under a TIL scenario using the Molecular Response to Immune Checkpoint Inhibitors (ICI) dataset (Eddy et al., 2020). |
| Dataset Splits | Yes | i CIFAR-10, i CIFAR100, and S-Tiny Image Net contain 3-channel images of size 32 32, 32 32 and 64 64 from 10, 100 and 200 classes, and each class includes 5000, 500, 500 training images and 500, 50, 50 test images. Each group of data is randomly split into a training set consisting of 80% of the samples and a test set containing the rest |
| Hardware Specification | Yes | All examples are executed on our machine with an AMD Ryzen7 2700 CPU and NVIDIA RTX 2060 GPU. |
| Software Dependencies | No | The paper mentions using a Resnet18 as a feature extractor, but does not provide specific version numbers for software dependencies like programming languages (e.g., Python) or libraries (e.g., PyTorch, TensorFlow, CUDA). |
| Experiment Setup | Yes | Our CL-BRUNO is specified as follows: We use a Resnet18 (He et al., 2016) as a feature extractor... In this example, we set the number of coupling layers in CL-BRUNO to be 6, the dimension of class embedding to be 128, size of pseudo data N = 128 (i.e. N pseudo samples are generated to compute the distributional regulariser in Eq 6 or 8 for every gradient descent step) and regularisation strength α1 = α2 = 1. For all experience replay-based methods, we fix the buffer size M = 500. |