# Attention-based Deep Multiple Instance Learning

**Authors:** Maximilian Ilse, Jakub M. Tomczak, Max Welling  
**arXiv:** [1802.04712](https://arxiv.org/abs/1802.04712)  
**Date:** 2018-02-13  
**Summary page:** https://www.summarizepaper.com/en/paper/1802.04712v1/

## Key points

- Maximilian Ilse, Jakub M. Tomczak, and Max Welling explore multiple instance learning (MIL), a supervised approach assigning a single class label to a bag of instances.
- The authors focus on formulating the MIL problem as learning the Bernoulli distribution of the bag label using neural networks.
- They propose a neural network-based permutation-invariant aggregation operator with an attention mechanism to understand individual instance contributions to the overall bag label.
- Empirical validation shows their approach achieves comparable performance to leading methods on benchmark datasets and outperforms others on various datasets, including MNIST and histopathology datasets.
- The research emphasizes interpretability without sacrificing accuracy by leveraging attention mechanisms in neural networks for MIL tasks.
- This work demonstrates how understanding individual instances' influence can enhance model performance and provide insights into decision-making processes in complex real-world applications.

## Summary

In their paper, "Attention-based Deep Multiple Instance Learning," Maximilian Ilse, Jakub M. Tomczak, and Max Welling explore the field of multiple instance learning (MIL), a supervised learning approach where a single class label is assigned to a bag of instances. The authors' main focus is on formulating the MIL problem as learning the Bernoulli distribution of the bag label, using neural networks to fully parameterize the bag label probability. One significant contribution of this research is their proposal of a neural network-based permutation-invariant aggregation operator that incorporates an attention mechanism. This innovative operator allows for understanding the individual contributions of each instance towards determining the overall bag label. Through empirical validation, it was shown that this novel approach not only achieves comparable performance to leading MIL methods on standard benchmark datasets but also outperforms other techniques on a MNIST-based MIL dataset and two real-life histopathology datasets. The significance of this work goes beyond just performance metrics; it also highlights interpretability without sacrificing accuracy. By leveraging attention mechanisms within neural networks for MIL tasks, the authors demonstrate how understanding the influence of individual instances can improve model performance and provide valuable insights into decision-making processes. Overall, "Attention-based Deep Multiple Instance Learning" presents a promising direction for advancing MIL methodologies with a focus on both effectiveness and interpretability in complex real-world applications.