https://www.overleaf.com/6277751678wbdzgcjpxcrf

Multi-instance (MI) classification is a subtype of supervised learning. Instead of working with individually labelled data points (instances), we work with bags (sets of instances). Class labels of bags are related to the instances within it using one of a range of possible assumptions.

For instance, under the standard assumption we view any bag with a positively labelled instance as positive itself. A good use case is the popular Musk dataset. This dataset comprises features describing molecules, some of which are classified as musks (a class of substances with a specific smell that have chemical properties useful in making a particular type of drug). However, each of these molecules appear in different conformations (shapes), which are pooled together. Therefore, a positive label is assigned to a set (bag) of all present conformations of a given molecule based on at least one of the conformations being classified as musk.

More elaborate assumptions include a looser threshold-based rule and assume a minimum number of positive instances required for a positive bag class label, or the collective assumption which views bags as probability distributions and lets us learn a classifier that can tell whether an instance comes from a positive or negative bag. Some MI learning algorithms label bags directly (eg. nearest-neighbour classifiers) while others do so through learning a concept for labelling instances individually (eg. neural networks adapted for a multi-instance setting).

An exciting use case for MI classification is image analysis. Taking patches (regions) of images as bags, we can classify entire images based on learning instance-level representations and combining them. Recent advances in MI learning include neural networks with residual connections (similarly as in ResNet CNN networks for image classification). Finally, the use of attention maps [1,2] can reveal which patches are responsible for a given bag class label. As an alternative to convolutional neural networks, this application holds promise in medical diagnostics due to good interpretability.

In this project, we ask you to study existing MI classification methods, up to the recent ones, and to present a use case on real-world data (eg. for drug-activity prediction, text categorisation or content-based image classification). Compare different classification algorithms and their parameters. Include a multiple-instance neural network in your comparison and evaluate its performance. We will appreciate if you explore the concept of attention weights and interpretability in your model.

There are numerous blog posts and lectures online that introduce multi-instance learning. The lecture part of this workshop gives a nice overview of both older and more recent concepts in the field. Skimming through Carbonneau et al. (2018) gives a good idea of what issues one has to tackle with different types of data (they have a GitHub repo as well, however they use MATLAB). Wang et al. (2018) give more information on MI neural networks, and provide TensorFlow implementations. Xiong et al. (2021) use data about T-cell receptors (structures active in the immune system) for cancer detection, comparing various multi-instance learning techniques. Datasets Here are some examples. The Musk dataset has been seminal in the field, and you can start off by testing your code on it. Similarly, for image data the MNIST handwritten digits dataset or CIFAR-10 are classics. In a next step, biomedical datasets like Skin Cancer MNIST: HAM1000 or chest X-ray images for pneumonia detection could be interesting to you, since they present an opportunity to use more advanced, deep-learning approaches to classification.

[1] Ilse, Maximilian, Tomczak, Jakub M., and Welling, Max. "Attention-based Deep Multiple Instance Learning". ICML. PMLR, 2018. -> https://github.com/garydoranjr/misvm -> https://github.com/AMLab-Amsterdam/AttentionDeepMIL

[2] Papadopoulos, Alexandros, Topouzis, Fotis, and Delopoulos, Anastasios. "An interpretable multiple-instance approach for the detection of referable diabetic retinopathy in fundus images". Nature Sci Rep, 2021.

[3] Carbonneau, Marc-André, et al. "Multiple instance learning: A survey of problem characteristics and applications". Pattern Recognition 2018.

[4] Wang, Xinggang, et al. "Revisiting multiple instance neural networks". Pattern Recognition, 2018. -> https://github.com/keras-team/keras

[5] Xiong, Danyi, et al. "A comparative study of multiple instance learning methods for cancer detection using T-cell receptor sequences". Computational and Structural Biotechnology Journal, 2021.