ZeroMamba: Exploring Visual State Space Model for Zero-Shot Learning

Authors: Wenjin Hou, Dingjie Fu, Kun Li, Shiming Chen, Hehe Fan, Yi Yang

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experiments on four prominent ZSL benchmarks, Zero Mamba demonstrates superior performance, significantly outperforming the state-of-the-art (i.e., CNN-based and Vi T-based) methods under both conventional ZSL (CZSL) and generalized ZSL (GZSL) settings.
Researcher Affiliation Academia 1Re LER Lab, Zhejiang University, China 2Huazhong University of Science and Technology (HUST), China 3Mohamed bin Zayed University of AI EMAIL
Pseudocode No The paper describes the proposed method using mathematical formulations (Eq 1-14) and block diagrams (Fig 2) but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Code https://github.com/Houwenjin/Zero Mamba
Open Datasets Yes To evaluate the effectiveness of our proposed framework, we conduct extensive experiments on three prominent ZSL datasets: Caltech-USCD Birds-200-2011 (CUB) (Welinder et al. 2010), SUN Attribute (SUN) (Patterson and Hays 2012) and Animals with Attributes 2 (AWA2) (Xian et al. 2019). We use the Proposed Split (PS) (Xian et al. 2019) division, as detailed in Tab. 3. Additionally, we verify the generalization of Zero Mamba on large-scale Image Net (Russakovsky et al. 2015) benchmark.
Dataset Splits Yes We use the Proposed Split (PS) (Xian et al. 2019) division, as detailed in Tab. 3. [...] We randomly split the training/test set into 800/200. We only use 1/10 of the data for every class for training.
Hardware Specification Yes We implement our experiments in Py Torch and utilize the SGD optimizer (momentum = 0.9, weight decay = 0.001) with learning rate of 5 10 4 on a single NVIDIA A100 GPU.
Software Dependencies No The paper mentions "Py Torch" but does not specify a version number, preventing reproducible software dependency.
Experiment Setup Yes We implement our experiments in Py Torch and utilize the SGD optimizer (momentum = 0.9, weight decay = 0.001) with learning rate of 5 10 4 on a single NVIDIA A100 GPU. All models are trained with a batch size of 16. In our method, we empirically set {λsc, λcol} to {1.0,0.3}, {0.2,0.35}, and {0.0,0.98} for CUB, SUN, and AWA2, respectively.