Are Multiple Instance Learning Algorithms Learnable for Instances?
Authors: Jaeseok Jang, HYUK-YOON KWON
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we conduct the following experimental validations to demonstrate whether existing Deep MIL approaches are learnable for instances based on the theorems: 1) (Theorem 5): Demonstrating the learnability of the attention mechanism for bags in DGen XY . 2) (Theorem 8, 9): Showing that multiplying attention at the feature level is not learnable for instances. 3) (Theorem 10, 11): Demonstrating that inputting position-related values into instance positions is not learnable for instances. As this study assumes an environment where bags are PAC Learnable, we preprocess the MNIST dataset to match the difficulty level of each experiment. For the validation of Theorem 10 and 11, we use the Web Traffic dataset from Early et al. [10], which is a synthetic time-series classification dataset. Detailed experimental settings can be found in Appendix D. |
| Researcher Affiliation | Academia | Graduate School of Data Science, Seoul National University of Science and Technology {jangjs1027, hyukyoon.kwon}@seoultech.ac.kr |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | In this study, to validate several final Theorems derived from the initial Theorems, we created appropriate experimental environments for each Theorem and conducted experiments. The results from these experiments provided empirical validation for the Theorems. ... In this study, experiments were conducted using the open datasets MNIST and the Web Traffic dataset [10]. For MNIST, since it was used to create synthetic datasets tailored to the experimental environment, we submit the code for data preprocessing and model training. For the Web Traffic dataset [10], as experiments were conducted by swapping models in the code published by Early et al. [10], we submit the code related to the model structures. |
| Open Datasets | Yes | In this study, experiments were conducted using the open datasets MNIST and the Web Traffic dataset [10]. |
| Dataset Splits | No | The paper mentions splitting the MNIST dataset into training and testing (8:2 ratio) but does not explicitly provide details for a separate validation split for all experiments. While it mentions 'the validation is adjusted to fit the nature of MIL' for MIREL, this does not constitute a general validation split specification for reproducibility across all experiments. |
| Hardware Specification | Yes | All experiments conducted in this study were performed on an Intel(R) Xeon(R) Silver 4210R 40 Core CPU @ 2.40 GHz, 32GB RAM, and NVIDIA RTX A5000. |
| Software Dependencies | No | The paper mentions specific MIL algorithms and states that some were implemented using 'official code provided by each respective paper', but it does not specify software dependencies like Python, PyTorch, or CUDA versions. |
| Experiment Setup | Yes | The hyperparameter settings for the models used in the experiments are shown in Table D. (Table 7: Optimizer: Adam, Learning rate: 0.0001, Cost function: NLL loss function, Epochs: 20). |