Tensorized Projection for High-Dimensional Binary Embedding
Authors: Weixiang Hong, Jingjing Meng, Junsong Yuan
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental comparisons with state-of-the-art methods over various visual tasks demonstrate both the efficiency and performance advantages of our proposed TP... Extensive experiments show that our approach not only achieves competitive performance in compact-bit case but also outperforms state-of-the-art methods in long-bit scenario. To evaluate Tensorized Projection (TP), we conduct experiments on three tasks... |
| Researcher Affiliation | Academia | Weixiang Hong, Jingjing Meng, Junsong Yuan School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore {wxhong,jingjing.meng,JSYUAN}@ntu.edu.sg |
| Pseudocode | No | The paper describes the optimization steps and the Alternating Least Square (ALS) method with mathematical equations and descriptive text but does not include a clearly labeled pseudocode block or algorithm box. |
| Open Source Code | No | The paper does not provide any statement about open-sourcing the code for its methodology or a link to a code repository. |
| Open Datasets | Yes | We use the pre-trained Alex Net (Krizhevsky, Sutskever, and Hinton 2012) provided by Caffe (Jia et al. 2014) to extract DNN features for one million images in MIRFLICKR-1M dataset (Huiskes and Lew 2008). We refer to this dataset as DNN4096. GIST-960 (Jegou, Douze, and Schmid 2008) dataset, which contains one million 960-dimensional GIST features (Oliva and Torralba 2001) and 10, 000 queries. Holidays + MIRFlickr-1M dataset (Jegou, Douze, and Schmid 2008). CIFAR-10 dataset (Krizhevsky 2009). |
| Dataset Splits | Yes | We use the pre-trained Alex Net (Krizhevsky, Sutskever, and Hinton 2012)... to extract DNN features for one million images in MIRFLICKR-1M dataset (Huiskes and Lew 2008). We randomly pick 1,000 samples as queries. We first fine-tune the pre-trained model provided by Caffe (Jia et al. 2014) on the training set of CIFAR-10, then we use the fine-tuned model to generate features for both training images and testing images. |
| Hardware Specification | Yes | The machine we use is equipped with Intel Xeon CPUs E5-2630 (2.30GHz) and 96 GB memory. We gratefully acknowledge the support of NVAITC (NVIDIA AI Technology Centre) for their donation of a Tesla K80 and M60 GPU used for our research at the ROSE Lab. |
| Software Dependencies | No | All experiments are conducted using Matlab, while the evaluation of encoding time is implemented in C++ using a single thread. The paper does not provide specific version numbers for Matlab, C++, or any other libraries used. |
| Experiment Setup | Yes | For the hyperparameter β in Equation 9, we experimentally find that a fixed β = 1 leads to competitive accuracy, and the accuracy is insensitive to the choice of β (we tried from 0.1 to 100). So we simply fix β = 1 for all experiments in this work. In practice, it takes around 10 iterations to converge as shown in Figure 1. Based on these observations, we fix the TT-rank to 4 to balance the accuracy and efficiency in all following experiments. |