A survey report on Multi-Instance Learning (MIL), written for COMP8230 at Macquarie University (a unit on mining unstructured data). The report explains what MIL is, how it is set up mathematically, the main families of methods, and where it gets used, with a closer read of object detection in computer vision.
The full report is in this repo as a PDF. This README is the short version.
- MIL learns from labels attached to groups of data (called bags) instead of every individual item, which cuts the cost of labelling.
- A bag counts as positive if at least one item inside it is positive. That single rule is the standard MIL assumption.
- It suits data that already comes in groups: an image is a set of regions, a molecule is a set of atoms and shapes, a document is a set of sentences.
- The trade-off is weaker supervision. You get cheaper labels, but you lose easy access to which item actually mattered, and the model is more prone to overfitting.
Standard supervised learning wants a label for every training example. To train a model to find a tumour in a scan, the usual approach expects someone to mark the exact region. That kind of labelling is slow, costly, and often not feasible at the level of detail you would want.
MIL changes what the label describes. Instead of labelling each piece, you label a group. The group is a bag and the pieces inside it are instances. You tell the model whether the whole bag is positive or negative, and nothing about the instances directly.
The usual way to picture it is a handful of keychains. Each keychain is a bag and each key is an instance. You know whether a given keychain can open a door, but not which key does it. A keychain that opens the door is a positive bag, and one working key is enough to make it positive. That is the standard MIL assumption in plain terms.
A lot of real data arrives in bags whether we plan for it or not, and the label that matters sits at the group level while the detail underneath stays messy or unlabelled. The report covers three settings where this happens:
- Medical imaging, where you can often say a scan contains a disease without marking exactly where it is.
- Drug discovery, where a molecule is tagged active or inactive depending on whether any of its 3D shapes can bind to a target site.
- Traffic and video systems, where objects appear and change from frame to frame and clean per-object labels are hard to maintain.
MIL fits all three because the honest label you can get is a group label, not a per-item one.
A bag X has a label y that is either 0 (negative) or 1 (positive). The model learns a function F(X) that returns a value in [0,1] for a bag it has not seen. Training uses a set of bags with their bag-level labels and nothing finer.
Because the model never sees which instance is responsible for a positive bag, a few problems follow:
- Label ambiguity. In a positive bag, any instance could be the reason.
- No instance-level answer by default. Often you do want to know which region or which molecule shape mattered, and the basic setup will not tell you.
- Overfitting. With weaker supervision, the model can pick up bag-level shortcuts instead of learning what makes an instance positive.
The report notes the common fixes for the instance problem: attention-based MIL, max pooling, and instance-level classifiers.
Here is the pipeline the report uses to contrast ordinary supervised learning with MIL:
flowchart LR
O["Object such as an image"] --> B["Bag"]
B --> I1["Instance 1"]
B --> I2["Instance 2"]
B --> IN["Instance N"]
I1 --> M["MIL model"]
I2 --> M
IN --> M
M --> L["Bag label: 0 or 1"]
The report uses the taxonomy from Amores (2013), which sorts MIL methods into three groups by where they do their work.
| Family | How it treats the data | Main assumption | Strengths and limits |
|---|---|---|---|
| Instance-based (e.g. Wrapper) | Trains an ordinary classifier on instances, where each instance borrows its bag's label, then combines instance predictions into a bag prediction | Standard MIL assumption | Cheap to run and reuses well-tuned single-instance algorithms; reasonably interpretable; can suffer from label noise because instances are mislabelled during training |
| Bag-based (e.g. bag-level SVM) | Compares whole bags using a similarity or kernel and classifies at the bag level | Collective assumption (all instances contribute to the bag label) | Scalable and effective when instance labels are noisy or irrelevant; little insight at the instance level |
| Embedded-space | Maps each bag into a single fixed-length vector, then applies a standard classifier | Either, depending on the mapping | Turns MIL into an ordinary supervised problem |
The split between the standard and collective assumptions is the thing to watch. The standard assumption says one positive instance is enough to make a bag positive. The collective assumption says every instance pitches in. Which one you pick changes how you model the problem and how much you can read off individual instances afterward.
The deep-dive section is object detection in computer vision. Here a bag is an image, an instance is a region within it, and the bag label says whether the target object shows up anywhere in the image. That framing has the exact ambiguity MIL handles well: you know the object is somewhere in the frame, you do not have clean labels for each region, and in video the scene keeps changing.
The papers the report reviews:
- Weakly Supervised Deep Detection Networks (WSDDN). Takes a CNN pretrained on large-scale image classification and fine-tunes it on a target set using only image-level labels. Evaluated on ImageNet ILSVRC 2012.
- MIL boosting. Trains a sequence of weak learners on instances and combines them into one strong classifier, adapting AdaBoost and gradient boosting to the MIL setting.
- MIL-SERBoost (2010). Combines MIL, semi-supervised learning, and online boosting to track objects in video as their appearance changes, and stays stable under varying lighting and motion.
- A particle-filter tracker (2011). Pairs MIL with a particle filter to keep several hypotheses about where the object is, tested on the CAVIAR dataset.
The report ends on the open problems rather than a tidy win. Instance-level interpretation is still hard, weak supervision leaves room for overfitting, and label ambiguity never fully disappears. It points at a few directions worth watching: better instance-level interpretability, pairing MIL with self-supervised learning, and methods that hold up in fast-changing real-world data.
- The full survey report (PDF)
- This README
Written for COMP8230, Semester 1 2025, School of Computing, Macquarie University.