Attention-based Deep Multiple Instance Learningの発表資料

presentation of Attention-based Deep Multiple Instance Learning

1. Attention-based Deep Multiple Instance Learning University of Yamanashi 6th-year student Taro Nitta

2. Multiple instance learning (MIL) a bag of instances + a single class label Predicting a bag label Discovering key instances The usability of current MIL model is open to question Interpretability Instance level accuracy is low

3. Contents modeling the bag label probability a transformation of instances to a low-dimensional embedding a permutation-invariant aggregation function a final transformation to the bag probab ility parameterizing all transformations using NN trainable weighted average operation helping to find key instances

4. MIL problem formulation Input : target variable : is a binary label assigned to a bag of instances and each instance has an individual label then, or

5. The bag label is distributed according to the Bernoulli distribution with the parameter This is a permutation-invariant procedure because we assume neither ordering nor dependency of instances in a bag. Symmetric function Assumption

6. Theorem1 A scoring function for a set of instances is a symmetric function, if and only if it can be decomposed in the following form: where are suitable transformation.

7. Theorem2 For any , a Hausdorff continuous symmetric function can be arbitrarily approximated by a function in the form are continuous functions.

8. Approach for classifying a bag of instances MIL pooling

9. Two main MIL approaches score MIL pooling instance embedding MIL pooling bag representation

10. MIL with Neural Networks MIL pooling MIL pooling MIL pooling operations max mean log-sum-exp Integrated Segm entation and Recognition noisy-or noisy-and

11. Attention-based MIL pooling instance-based embedding-based

12. A ttention-based MIL pooling

13. A ttention-based MIL pooling

14. FC layer FC layer FC layer Softmax sigmoid gate