Adaptive Few-Shot Class-Incremental Learning via Latent Variable Models
Authors: Tameem Adel
JAIR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on several widely used few-shot class-incremental learning benchmarks, as well as a medical benchmark consisting of real-world medical images, demonstrate that the proposed model leads to improved performance, as measured by average overall and final classification accuracy, and in terms of alleviating catastrophic forgetting. |
| Researcher Affiliation | Academia | Tameem Adel EMAIL National Physical Laboratory, Maxwell Centre, University of Cambridge JJ Thomson Avenue, Cambridge, CB3 0HE, United Kingdom |
| Pseudocode | Yes | Algorithm 1 The proposed CIAM algorithm for few-shot class-incremental learning adaptation via latent variable modeling |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code for the methodology, nor does it provide a link to a code repository. |
| Open Datasets | Yes | We evaluate CIAM here by performing experiments on the following FSCIL benchmarks: mini Image Net (Russakovsky et al., 2015), CIFAR100 (Krizhevsky & Hinton, 2009), and CUB200 (Wah et al., 2011). ... In addition to the commonly used FSCIL benchmarks, we apply the proposed CIAM framework to real-world medical images in the form of the Med MNIST dataset (Yang et al., 2021a, 2021b). |
| Dataset Splits | Yes | mini Image Net: ... The base session, t = 1, comprises 60 classes. After the base session, there are 8 few-shot sessions, each comprising 5 classes. ... each few-shot session, 2 <= t <= 9, is a 5-way 5-shot session, which means that there are 5 training points available for each of the 5 classes. CIFAR100: ... beginning with a base session of 60 classes, followed by 8 5-way 5-shot sessions. CUB200: ... the 200 classes are divided into 100 classes for the base session, followed by 10 incremental 10-way 5-shot sessions. Med MNIST: ... 6 selected medical disease classification datasets. The classes involved in 3 (out of the 6) medical datasets are selected as the classes of the base session (with a total of 27 classes), whereas the classes of the other 3 datasets serve as the incremental few-shot classes (with a total of 15 classes). ... We proceed with 15 1-way 1-shot incremental few-shot sessions each containing one out of the 15 few-shot classes. |
| Hardware Specification | Yes | The computational environment consists of an NVIDIA A100 Tensor Core GPU and two AMD Rome CPUs based on the NVIDIA Mellanox Connect X-6 interconnect technology. |
| Software Dependencies | No | The paper mentions using "stochastic gradient descent (SGD) with momentum" as an optimizer, but it does not specify any software libraries or packages with version numbers. It refers to "Res Net-18 (He et al., 2016)" as the backbone network, which is a model architecture, not a specific software dependency with a version number. |
| Experiment Setup | Yes | As an optimizer, we use stochastic gradient descent (SGD) with momentum, with an initial learning rate of 0.01 for both mini Image Net and CIFAR100, and an initial learning rate of 0.001 for the CUB200 dataset. The kernel width parameter σ is tuned using cross-validation. The number of normalizing flow steps s is 2... we train the model on all the experiments with a minibatch size of 512 during the base session, and a minibatch size of 64 during each incremental few-shot session. For the mini Image Net dataset, the number of epochs is 500 for the base session, and is 150 for each incremental few-shot session. For CIFAR100, we train the model for 200 epochs during the base session and for 100 epochs during each few-shot session. The number of epochs for the CUB200 dataset is 80 epochs for the base session and 60 epochs for each incremental few-shot session. |