Machine Learning

Supervisors: Faizan Ahmed

Point clouds are sets of spatial data points captured by 3D scanning techniques such as lidar. These point clouds contain many million points of data, resulting in 3D representations of the railway environment. Point cloud data can be used to create machine learning models that can classify the object in rail infrastructure automatically. 

The main research question is: How to reliably segment, classify, and visualize point clouds for railway catenary systems with scant computational power, memory, and ground truth labels?

However, the student can focus on a (combination of) topic(s) given below:

• Create an end-to-end pipeline for point cloud visualisation using existing libraries
• visualizing model output(or intermediate processing) to explore interpretability/explainability
• application of XAI techniques to point clouds
• Leverage existing trained models for point cloud labelling
• pre-process (down-sampling, Augmentation, filtering, voxelization etc.) point clouds for accurate model training
• train deep learning models for point cloud segmentation and object detection

Some previous assignments related to this topic:

• Verburg, F.M. (2022) Exploring explainability and robustness of point cloud segmentation deep learning model by visualization.
• Vieth, Z.J. (2022) Point cloud classification and segmentation of catenary systems.

In Data Science, much effort goes into the implementation of a data preprocessing and model training pipeline. It is important that the resulting system is validated on many more aspects than just the predictive power of the resulting model. One aspect that is in dire need of additional validation is the data preprocessing part. It should produce data of sufficient quality for model training and testing. The main question in this research project is: how to effectively validate a data preprocessing pipeline?

Some of this validation should already be done before all implementation has been completed, i.e., by specifying of the preprocessing pipeline on a high level, and checking whether basic properties hold. After implementation, the specification can be used for model-based testing. Here the high-level specification models what the pipeline is expected to do. Specifically, in model-based testing the two following steps are taken: (i) a test generation algorithm is applied on the specification to automatically derive tests, and (ii) these tests can then be executed on the implementation of the pipeline to check that it functions as specified.

The idea for the project is to establish a high-level modeling language for data preprocessing pipelines, and an approach, based on Model-Based Testing, that can then validate this data preprocessing pipeline.

Example: Imagine a data science project developing a machine learning model for predicting the duration of heart surgeries for the purpose of improving the planning of surgeries. The developed predictive model is meant to predict the duration of a surgery given data on the characteristics of the patient, their condition, the type of surgery, the staff involved, and contextual factors. The data preprocessing will include many cleaning actions of individual attributes, deletion of rows with too many missing values, derive some other attributes, etc. One could imagine making a high-level specification of these steps which can then be validated on properties like completeness (is still a minimum of 80% of the data left?) and bias (are still certain subsets of surgeries represented in the data with roughly the same ratios?)

Supervisor: Faizan Ahmed

Black-box machine learning models are frowned upon by regulators. Therefore, it is essential to develop methods and tools to peek inside these black boxes. Explainable AI (XAI) is a set of methods and tools for explaining or reasoning about the outcomes or internal workings of these black-box models.

Explainable AI has many different facets, which lead to several interesting research directions. This topic is therefore not narrowed down to one specific research question; instead, it is presented as a set of research directions that I find interesting. A few concrete directions are as follows:

1. Explainable AI for time series or sequence models: The focus could be on developing new XAI techniques or adapting existing ones to the time series data modality. A very concrete project in this direction is a comparative study of the TCAV and C-SHAP frameworks.
2. Explainable AI for point clouds: Although point cloud analysis is formally a computer vision task, the data modality requires specialist tools and techniques, since its data structure differs substantially from that of images.
3. Explanation multiplicity: If two models differ only in their hyperparameters, what can be said about the relationship between the explanations they produce? This is a new and emerging area within explainable AI.
4. Explanation by or for LLMs: LLMs offer a way to present explanations in a human-understandable form. On the other hand, LLMs are themselves black-box models that require explanation. In this regard, mechanistic interpretability is an interesting direction to explore.
5. Applied explainable AI: Applying existing explainability techniques in domains such as healthcare and finance comes with its own challenges, so this could be a standalone research topic in itself.

If you find any of the above directions interesting, or have a project of your own in explainable AI, feel free to contact me.