Introduction
Multiple Instance Learning (MIL) involves assigning labels to "bags" of instances rather than individual items. A classic example is medical imaging, where a whole slide is labeled "cancerous" if it contains at least one malignant patch. Traditionally, MIL architectures have focused on permutation-invariant pooling methods, assuming that Long Short-Term Memory (LSTM) networks—designed for sequential data—are unsuitable for unordered bags. This paper challenges this assumption, demonstrating that LSTMs, when trained with random sequence augmentation, can outperform specialized MIL models by leveraging their inherent memory mechanisms.
Methodology
The proposed framework consists of four primary components:
- Instance Description Unit (IDU): A feature extractor (e.g., LeNet or VGG) that converts raw instances into vector representations.
- Iterative Bag Pooling Unit: The core of the architecture, utilizing a Bi-directional LSTM to process instances sequentially. To ensure permutation invariance, instances are fed in a random order during training.
- Mutual Information Maximization (MIM): An auxiliary loss function designed to maximize the information captured by the learned representations from the input instances.
- Prediction Unit: A final layer that maps the LSTM's terminal hidden state to a bag-level prediction.
The model can also perform instance-level classification by treating individual instances as "singleton bags" and analyzing the resulting internal states.
Experimental Setup
The approach was evaluated across three domains:
- MNIST-MIL: Synthetic benchmarks including single/multiple digit occurrence, digit counting, and outlier detection.
- Lookbook: A fashion dataset for cross-domain clothing retrieval, including an "occluded" variant where bags contain image fragments.
- Histopathology (Colon Cancer): A medical dataset requiring the identification of cancerous tissue within large images composed of thousands of cell nuclei.
Results and Discussion
The experimental results indicate that LSTMs excel in "collective" MIL problems where labels depend on complex relationships between multiple instances. On the MNIST "Multiple Digit Occurrence" task, the LSTM-based model outperformed state-of-the-art attention-based methods by approximately 27 percentage points.
Key insights include:
- Memory over Order: The LSTM's gating mechanism effectively "accumulates and remembers" trigger instances (e.g., cancerous cells) regardless of their position in the sequence.
- Robustness: Shuffling bag order at test time has negligible impact on accuracy, confirming that random order training successfully induces permutation invariance.
- Weakly Supervised Success: On the Colon Cancer dataset, the model achieved 92.74% bag-level accuracy and 82.9% instance-level accuracy, the latter being remarkably close to fully supervised performance (85.0%).
Conclusion
LSTMs are surprisingly effective for unordered MIL tasks, particularly those requiring collective reasoning across instances. While performance is sensitive to significant discrepancies between training and testing bag sizes (cardinality), this can be mitigated through fine-tuning. Future research will address computational efficiency for extremely large bags where linear processing may become a bottleneck.
Summary updated