Principles of AI-driven Neuroscience and Explainability

Aims

This module aims to provide students with a comprehensive understanding of graph neural networks, selected geometric deep learning models, convolutional neural networks (CNNs), transformers, and fundamental concepts of brain anatomy, connectomes, and medical imaging. A particular emphasis is placed on explainability and interpretability in the medical domain, specifically neuroscience.
By the end of the module, students will be familiar with classification tasks involving practical examples of brain pathologies, such as psychotic, brain cancer and neurodegenerative disorders. They will also explore how AI can potentially drive neuroscience forward by identifying novel patterns and enhancing our understanding of brain function and anatomy.
The module combines lectures and some hands-on applications to ensure both theoretical depth and practical experience covering foundational topics such as:

1. Introduction to Basic Brain Anatomy, Connectomes and Medical Imaging
2. Introduction to Basic Deep Learning and Classification Structures in Neuroscience
3. Graphs Neural Networks and Neuroscience
4. Geometric Deep Learning Applied in Neuroscience
5. Principles of Attributional Interpretability Methodologies
6. Principles of Mechanistic Interpretability Approaches
7. Applications of Explainable Artificial Intelligence in Neuroscience
8. Clinical neuroscience perspective

Syllabus

Lecture 1: Introduction to Basic Brain Anatomy, Connectomes and Medical Imaging
The aim of this lecture is to introduce basic brain anatomy, focusing on different lobes, sulci regions, and the brain’s folding patterns. Additionally, it will introduce the fundamental characteristics of various medical imaging modalities, with an emphasis on anatomical and functional MRI. The lecture will also present methods for extracting connectome information [25] using imaging techniques such as functional and structural MRI. [1,2]

Lecture 2: Introduction to Basic Deep Learning and Classification Structures in Neuroscience
This lecture will briefly recap the basics of deep learning architectures and layers, such as MLPs, transformers, attention layers, and CNN blocks. It will also present state-of-the-art classifier architectures like ViT and introduce classification problems in neuroscience. [3,4]

Lecture 3: Graphs Neural Networks and Neuroscience
This lecture will introduce the fundamental concepts of graph representation in neuroscience, focusing on how connectomics [21] can be used to model brain imaging data by using nodes as brain regions and edges as connections. Additionally, we will discuss Graph Neurla Networks (GNNs), covering essential architectures such as GCNs [22] and GATs [23].

Lecture 4: Geometric Deep learning Applied in Neuroscience
This lecture will extend the previous concepts on graph neural networks by covering a basic introduction to geometric deep learning principles [6] and
hyperbolic space. The main aim is to discuss specific geometric deep learning architectures that benefit from the spherical representation of brain anatomy. These architectures include spherical CNNs [5], GNNs [19], hyperbolic GNNs [7], and hyperbolic CNNs [8].

Lecture 5: Principles of Attributional Interpretability Methodologies
This session will highlight various Explainable AI (XAI) techniques, focusing primarily on the principal methods used in attributional interpretability, such as LIME, LRP, GradCam, SHAP and GNNExplainer. [9-12, 24]

Lecture 6: Principles of Mechanistic Interpretability Approaches
This lecture will cover key terminology related to superposition, polysemantic representations, and the privileged basis. Additionally, it will address the mechanistic interpretability problem and explore how the sparse auto-encoder attempts to provide explanations for various deep learning applications. [13,14]

Lecture 7: Applications of Explainable Artificial Intelligence in Neuroscience This session will provide a brief application of all previous lectures, focusing on real-world implementations in neuroscience. It will cover tasks such as classification [15,20], XAI [16,17], and mechanistic interpretability [18]. Additionally, we will discuss the impact of AI in neuroscience and the importance of responsible AI practices that align with ethical considerations in medical applications. Lecture 8: Clinical neuroscience perspective This session will primarily focus on exploring real-world challenges and applications of AI in clinical practice, providing students with insights into the perspective of healthcare professionals. To enhance their understanding, we plan to invite an expert in the field, such as Professor Murray Graham or Professor John Suckling, to discuss the integration of AI in hospital settings and its impact on clinical decision-making.
 

Learning outcomes

This module is designed not only to equip students with technical expertise in AI-driven neuroscience but also to instil a strong ethical and philosophical perspective on AI's role in medical and societal contexts. Students will develop practical skills in applying state of the art AI networks and explainable AI (XAI) techniques to real-world problems in neuroscience, enabling them to critically assess and refine AI methodologies.
By the end of the module, students will be able to:

• Apply graph neural networks, geometric deep learning models, CNNs, and transformers to neuroscience-related classification tasks.
• Interpret AI-driven models in medical imaging and connectomes with a focus on explainability and trustworthiness.
• Basic analyse brain pathologies, such as psychotic, neurodegenerative disorders, and brain cancer, through AI-based classification techniques.
• Evaluate the impact of AI in neuroscience and engage in responsible AI practices that align with ethical considerations in medical applications.

Through a combination of lectures and hands-on applications, students will gain the necessary skills to contribute to the advancement of AI in neuroscience research. 

Assessment

- (85%) Group Mini-project (report writeup) at the end of the course. The projects will be developed in pairs, and the students can either self-propose their own ideas or express their preferences for one of the provided topics announced at the start of term. The writeup report will have a limit of 4,000 words in line with other modules. 

- (15%) Short presentation and viva. Students will deliver a brief presentation outlining their individual contributions to the mini project, followed by a short viva.

Assessment criteria for writeup report will follow project assessment criteria here: https://www.cl.cam.ac.uk/teaching/exams/acs_project_marking.pdf

Recommended reading material and resources

Theory: 

[1]https://www.sciencedirect.com/science/article/pii/S1053811913002656
[2] https://pmc.ncbi.nlm.nih.gov/articles/PMC1239902/ 
[3] https://arxiv.org/abs/2305.09880 
[4] https://arxiv.org/abs/2010.11929 
[5] https://arxiv.org/abs/1801.10130 
[6] https://arxiv.org/abs/2105.13926 
[7] https://arxiv.org/abs/1910.12892 
[8] https://arxiv.org/abs/2303.15919 
[9] https://arxiv.org/abs/1602.04938 
[10] https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0130140 
[11] https://arxiv.org/abs/1705.07874 
[12] https://arxiv.org/abs/1610.02391 
[13] https://arxiv.org/abs/2209.10652 
[14] https://arxiv.org/abs/2309.08600
[15] https://www.nature.com/articles/s43856-023-00313-w 
[16] https://arxiv.org/abs/2309.00903 
[17] https://arxiv.org/abs/2405.10008 
[18] https://www.biorxiv.org/content/10.1101/2024.11.14.623630v1 
[19] https://www.sciencedirect.com/science/article/pii/S1361841524000471 
[20] https://www.sciencedirect.com/science/article/pii/S1361841522001189
[21] https://www.sciencedirect.com/science/article/pii/S105381190901074X 
[22] https://arxiv.org/abs/1609.02907 
[23] https://arxiv.org/abs/1710.10903 
[24] https://arxiv.org/abs/1903.03894 
[25] 10.1016/j.neuroimage.2011.05.025

Books: 

[1] "Deep Learning", by Ian Goodfellow, Yoshua Bengio and Aaron Courville. 
[2] “Graph Representation Learning Book", by William L. Hamilton. 
[3] “Graph Neural Networks: Foundations, Frontiers, and Applications”, by Lingfei Wu, Peng Cui, Jian Pei, and Liang Zhao. 
[4] “Changing Connectomes: Evolution, Development, and Dynamics in Network Neuroscience.” by Kaiser, M. (2020). MIT Press. ISBN: 978-0262044615

Coding: 

[1] https://pytorch-geometric.readthedocs.io/en/latest/ 
[2] https://captum.ai/ 
[3] https://pytorch.org/hub/huggingface_pytorch-transformers/ 
[4] https://huggingface.co/ 
[5] https://transformerlensorg.github.io/TransformerLens/ 
[6] https://sites.google.com/site/bctnet/