Introduction
Multiple Instance Learning (MIL) is a paradigm where labels are assigned to sets (or "bags") of instances rather than individual data points. Traditionally, MIL assumes a bag is positive if it contains at least one positive instance. However, many real-world scenarios require more complex logic, such as identifying specific combinations of items or counting occurrences.
Existing methods often rely on "stateless" pooling layers like Max-pooling or Attention mechanisms. While efficient, these approaches struggle to maintain a memory of previously seen instances within a bag, limiting their ability to perform tasks like counting or outlier detection. This paper proposes a simple yet effective baseline using Recurrent Neural Networks (RNNs) to address these limitations.
Methodology
The proposed framework consists of three primary components designed to process sets as sequences while maintaining a robust internal state.
- Instance Description Unit (IDU): This unit employs a feature extractor (e.g., a CNN) to transform raw data into vector representations. To ensure these features are informative, the authors incorporate Mutual Information Maximization (MIM) as a regularizer.
- Iterative Set Pooling Unit (ISPU): At the heart of the model is a Bi-directional LSTM. Instances are fed into the LSTM in a random order. The hidden state serves as a "running memory," allowing the model to aggregate information iteratively and capture dependencies between instances.
- Prediction Unit: The final hidden state of the LSTM is mapped to a bag-level label for classification or regression tasks.
Furthermore, the model supports weakly supervised instance-level learning. By analyzing the representations of single instances through the trained network, the system can infer instance-level labels without explicit supervision.
Experimental Evaluation
The model was tested across several domains, demonstrating significant improvements over attention-based baselines:
- MNIST Synthetic Tasks: In outlier detection, the LSTM-based approach reduced the error rate from ~37% (attention-based) to just 2.07%. In digit counting tasks, it outperformed baselines by approximately 24 percentage points.
- Medical Imaging: On the Colon Cancer dataset, the model achieved 92.74% accuracy in identifying epithelial nuclei.
- Fashion Retrieval: Using the Lookbook dataset, the model successfully matched catalog images to human-worn items by modeling the distribution of clothing items within a set.
Discussion and Conclusion
The primary insight of this work is that LSTMs, despite being sequence models, are highly effective for set-based learning because they maintain a state. This allows for complex reasoning—such as counting—that stateless pooling layers cannot easily replicate. While processing instances sequentially introduces $O(N)$ complexity, the performance gains in structured MIL tasks suggest that memory-augmented pooling is a vital component for advanced weakly supervised learning.
Summary updated