MetaZSCIL: A Meta-Learning Approach for Generalized Zero-Shot Class Incremental Learning
Authors: Yanan Wu, Tengfei Liang, Songhe Feng, Yi Jin, Gengyu Lyu, Haojun Fei, Yang Wang
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments conducted on five widely used benchmarks demonstrate the superiority of our proposed method. |
| Researcher Affiliation | Collaboration | Yanan Wu1*, Tengfei Liang1*, Songhe Feng1 , Yi Jin1 , Gengyu Lyu2, Haojun Fei3, Yang Wang4 1Beijing Key Laboratory of Traffic Data Analysis and Mining School of Computer and Information Technology, Beijing Jiaotong University 2Faculty of Information Technology, Beijing University of Technology 3360 Digi Tech, Inc 4Department of Computer Science and Software Engineering, Concordia University {ynwu0510, tengfei.liang, shfeng, yjin}@bjtu.edu.cn, lyugengyu@bjut.edu.cn, zhangchulan-jk@360shuke.com, yang.wang@concordia.ca |
| Pseudocode | Yes | Algorithm 1: The optimization procedure of Meta ZSCIL |
| Open Source Code | No | The paper does not provide any specific repository links or explicit statements about the availability of the source code for the methodology described. |
| Open Datasets | Yes | We conduct experiments on five widely used ZSL datasets: Animals with Attributes 1&2 (AWA1 (Lampert, Nickisch, and Harmeling 2013) & AWA2 (Xian et al. 2018a)), UCSD Birds-200-2011 (CUB (Wah et al. 2011)), Scene Recognition (SUN (Patterson and Hays 2012)), and Attributes Pascal and Yahoo (APY(Farhadi et al. 2009)). |
| Dataset Splits | Yes | AWA1 includes 30,475 images and AWA2 consists of 37,322 images. They are split into 40 seen classes and 10 unseen classes; CUB contains 11,788 images of 200 bird species in which 150 classes are treated as seen and 50 classes are unseen; SUN consists of 14,340 fine-grained images from 717 classes, including 645 seen classes and 72 unseen classes. In a PY, there are 15,339 images belonging to 32 classes, and 20 of these classes are treated as seen and 12 are unseen. Specifically, at each task, we first randomly separate the seen classes as pseudo seen classes and pseudo unseen classes without overlapping. Next we sample a sequence of T + 1 sessions, Ds = {(Dj tr, Dj te)}T j=0, where Dj tr and Dj te are the training and test set for the jth session. |
| Hardware Specification | No | The paper does not specify any particular hardware (e.g., GPU models, CPU types, or cloud instances) used for running the experiments. |
| Software Dependencies | No | The paper mentions that Our networks are optimized by the Adam optimizer with β1 = 0.5, β2 = 0.999, and initial learning rate 0.0001 but does not provide specific software dependencies with version numbers, such as programming languages, libraries, or frameworks. |
| Experiment Setup | Yes | Our networks are optimized by the Adam optimizer with β1 = 0.5, β2 = 0.999, and initial learning rate 0.0001 in both offline training and online incremental learning stages. The penalty coefficient λ is set to 10. The input noise in the generator has the same dimension as the corresponding attributes. We set mini-batch to 512 for AWA1 and AWA2, 64 for CUB, SUN and a PY. In the meta-training (offline) stage, we first perform supervised training on pseudo seen classes for 30 epochs, followed by 5 inner and 1 outer gradient updates for adapting new pseudo unseen classes without forgetting. In the meta-testing (online) stage, we directly perform 10 gradient updates to fast adapt unseen classes of each incremental session. |