Query-Driven Multi-Instance Learning
Authors: Yen-Chi Hsu, Cheng-Yao Hong, Ming-Sui Lee, Tyng-Luh Liu4158-4165
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on action classification over video clips and three MIML datasets from MNIST, CIFAR10 and Scene are provided to demonstrate the effectiveness of our method. |
| Researcher Affiliation | Collaboration | Yen-Chi Hsu,1,3 Cheng-Yao Hong,1 Ming-Sui Lee,3 Tyng-Luh Liu1,2 1Institute of Information Science, Academia Sinica, 2Taiwan AI Labs 3Department of Computer Science & Information Engineering, National Taiwan University |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statements or links indicating that the source code for the methodology is publicly available. |
| Open Datasets | Yes | We evaluate our method mainly on the MNIST-based dataset (MNIST-BAGS) (Ilse, Tomczak, and Welling 2018) and CIFAR10-based dataset (CIFAR10-BAGS)... We follow the similar data sampling method in (Ilse, Tomczak, and Welling 2018) to create the MNIST-BAGS MIL dataset from MNIST (Le Cun, Cortes, and Burges 1998) and analogously from CIFAR10 (Krizhevsky and Hinton 2009)... For fair comparisons, we adopt the MIML Scene dataset (Zhou and Zhang 2007) as the benchmark... particularly, we explore the problem involving the Activity Net (Fabian Caba Heilbron and Niebles 2015). |
| Dataset Splits | Yes | The standard MIL problem with one single query proceeds as follows. In MNIST or in CIFAR10, each of the ten categories will be chosen in turn as the one of interest, and the remaining are treated as background/noise... for each single query to a specific class label we first sample 500 training bags, including 250 positive and 250 negative bags from MNIST. Analogously, another 1000 bags (500 + & 500 ) are also generated for testing... For fair comparisons, we adopt the MIML Scene dataset (Zhou and Zhang 2007) as the benchmark and report 10-fold crossvalidation results. |
| Hardware Specification | No | The paper mentions 'National Center for High-performance Computing for providing computational resources and facilities' but does not provide specific hardware details such as GPU/CPU models, memory, or cloud instance types used for experiments. |
| Software Dependencies | No | The paper mentions the use of 'Adam' for optimization and references 'Le Net architecture' and 'Res Net50', but it does not specify version numbers for any software libraries, frameworks, or programming languages used for implementation. |
| Experiment Setup | Yes | The learning rate is 10 4 at initialization and the optimization method is Adam (Kingma and Ba 2014). The weight decay is 10 5, while λ in (4) is 10 4 for all the experiments. We fix τ in (3) as 0.5. σ1 and σ2 in (2) are tanh and linear mapping. |