KALAHash: Knowledge-Anchored Low-Resource Adaptation for Deep Hashing
Authors: Shu Zhao, Tan Yu, Xiaoshuai Hao, Wenchao Ma, Vijaykrishnan Narayanan
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our proposed method, Knowledge Anchored Low-Resource Adaptation Hashing (KALAHash), significantly boosts retrieval performance and achieves a 4 data efficiency in low-resource scenarios. Table 1 presents the m AP results for NUS-WIDE, MSCOCO, and CIFAR-10 across different low-resource settings (1-8 shots). |
| Researcher Affiliation | Collaboration | Shu Zhao1, Tan Yu2, Xiaoshuai Hao3, Wenchao Ma1, Vijaykrishnan Narayanan1 1The Pennsylvania State University 2NVIDIA 3Samsung EMAIL |
| Pseudocode | No | The paper describes the CLo RA and KIDDO methodologies using textual descriptions and mathematical equations, but does not include explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code https://github.com/Tree-Shu-Zhao/KALAHash.pytorch |
| Open Datasets | Yes | We evaluate our proposed method on three standard benchmarks: NUS-WIDE (Chua et al. 2009), MS-COCO (Lin et al. 2014), and CIFAR-10 (Krizhevsky and Hinton 2009). |
| Dataset Splits | Yes | NUS-WIDE is a multi-label dataset... We randomly select 2, 100 images (1, 00 images per class) to form the query set, and the rest is used as the gallery set. MS-COCO is a multi-label dataset... We randomly choose 5, 000 images as the query set, and the rest are viewed as the gallery set. CIFAR-10 consists of 60, 000 images... we randomly sample 1, 000 images (100 images per class) to construct the query set and the rest is used to form the gallery set. In our low-rank adaptation setting, we randomly split NK samples per class to create the training set. NK is 1, 2, 4, or 8 in our experiments. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for running its experiments. It only generally refers to 'backbone models' and discusses inference times without hardware specifications. |
| Software Dependencies | No | The paper mentions using specific models (e.g., CLIP Vi T-B/32) and optimizers (SGD) but does not specify the versions of key software libraries or frameworks like Python, PyTorch, or TensorFlow that would be needed for reproducibility. |
| Experiment Setup | Yes | η and r are set to 1.0 and 1, respectively. We use a 16-bit hash layer as default. We freeze all the parameters expected for the CLo RA module, G, F, and hash layer. We use SGD with 0.9 momentum and 1e 5 weight decay as the optimizer. The learning rate is set to 0.01. The batch size is set to 8. α, β, and γ are set to 0.1, 1.0, and 3.0, respectively. |