Orchestrator
NLP input is parsed into intent, constraints, data requirements, and expected output before the orchestrator selects the right worker.
L205 ACS Master module
Principles of AI-Driven Neuroscience and Translational Biomedicine
A research-led module on how neuroscience can inspire new AI systems, and how modern AI can reveal structure, mechanism, and clinically meaningful patterns in brain, multi-omics, imaging, and language data.
8
two-hour lectures
20%,65%,15%
theoretical-test, mini-project, presentation
2
students per project
Course idea
From Neuroscience to interpretable AI, and back again.
The module connects brain anatomy, connectomes, neuroimaging, multi-omics, clinical big data, deep learning, graph learning, large language models, and translational applications. The aim is not only to predict outcomes, but to understand where predictive power comes from (explanation and explainable-AI).
Neuroscience
AI
Attributional explainability
Feature importance, saliency, Grad-CAM, SHAP, GNNExplainer, and biomarker validation for models working with neuroimaging, neurobiology, and medical language.
Mechanistic interpretability
Information flow inside models, sparse autoencoders, superposition, polysemantic features, and circuit-level explanations for medical LLMs and biology.
Memory and forgetting
Catastrophic forgetting, model collapse, continual learning, replay, LoRA, and neuroscience foundations of memory in human and artificial systems.
Preview
The success course story of last year.
A curated recap of the key scientific concepts from last year’s module.
Agentic-AI highlight
Brain-Inspired Agentic AI: From Orchestration to Understanding.
This paradigm presents how the design of AI agents can draw inspiration from neuroscience. Conversely, interpreting the roles of coordinating orchestrators and specialised worker agents may provide useful computational hypotheses for understanding brain function. These connections underscore the importance of the module and demonstrate how its concepts can be translated into broader research and practical applications.
The video starts with a schematic brain-to-agent map, then follows user language through an orchestrator, specialised workers, reasoning, XAI, validation, memory, and catastrophic forgetting.
Workers
Specialist workers retrieve evidence, run models, analyse outputs, and return reasoning traces that can be inspected.
Trust
Validity comes from task checks, benchmarks, regression tests, uncertainty, attributional XAI, and mechanistic interpretability.
Forgetting
Continual learning can damage older skills. Replay, regularisation, adapters, monitoring, and drift tests help mitigate catastrophic forgetting.
Brain analogy sources:
frontoparietal control
visual cortex
language network
working memory
numerical reasoning
hippocampal memory
Syllabus
Eight lectures across brain structure, biological scale, AI methods, and translation.
Brain Anatomy, Connectomes and Neuroimaging
Lobes, sulci, folding patterns, structural and functional MRI, and extracting connectome information from imaging data.
Multi-Omics and Neurobiology
Genomics, transcriptomics, proteomics, metabolomics, microbiome layers, cellular heterogeneity, and AI-driven integration.
Deep Learning and Geometric Deep Learning
MLPs, CNNs, transformers, attention, ViT, graph representations, GCNs, GATs, and connectomic graph learning.
Clinical Big-Data Neuroscience
Large clinical datasets, UK Biobank-style pipelines, population-level neuroscience, and clinically grounded discovery.
LLMs, Catastrophic Forgetting and Neuroscience
Continual learning, replay, regularization, HAT, gradient memory, LoRA, model collapse, and cognitive neuroscience parallels.
Attributional Interpretability
Grad-CAM, SHAP, GNNExplainer, saliency, feature importance, and explainability in m-LLMs, neuroimaging, and neurobiology.
Mechanistic Interpretability
Superposition, polysemanticity, privileged bases, sparse autoencoders, information flow, and biological model explanations.
Applications in Omics, Reports and Neuro-Imaging
Clinical case studies, translational decision support, invited research and industry perspectives, and integrated AI-neuroscience workflows.
Assessment
Structured assessment with theory, collaboration, and individual defence.
Theory test
A 30-minute, in-class, closed-book test after Lecture 5 and before Lecture 7, focused on the fundamental concepts introduced in Lectures 1 to 5.
Group mini-project report
A pair-based project with a 4,000-word report. Students may propose a project or select from structured topics provided at the start of term.
Presentation and viva
A short presentation on individual contributions, followed by a brief viva to discuss decisions, results, interpretation, and project ownership.
Project fairness is supported through report contribution statements, individual teammate contribution statements, two dedicated progress discussion sessions, and first-line PhD student support for project questions.
Learning outcomes
Students leave able to work across models, mechanisms, and biomedical data.
- Build a multi-scale view of neuroscience from molecules to systems.
- Apply deep learning and graph learning to brain and biomedical datasets.
- Evaluate AI in clinical contexts, including reliability and forgetting.
- Use attributional and mechanistic interpretability to validate patterns.
- Analyse translational case studies in psychosis, brain cancer, and neurodegeneration.
- Critically assess how AI can advance neuroscience and clinical decision support.
Reading and tools
A compact launchpad for theory, books, and coding frameworks.
Neuroimaging foundations
Brain connectivity
Transformers and modern AI
Graph attention networks
Graph convolutional networks
Grad-CAM
GNNExplainer and XAI
Mechanistic interpretability
Single-cell and omics AI
AI for biological discovery
Deep Learning, Ian Goodfellow, Yoshua Bengio and Aaron Courville
Graph Representation Learning, William L. Hamilton
Graph Neural Networks: Foundations, Frontiers, and Applications, Lingfei Wu et al.
Changing Connectomes, Marcus Kaiser
Principles of Neurodynamics, Frank Rosenblatt
Mind in Motion, Barbara Tversky
Cognitive Foundations of Agentic AI, Anand Vemula
Machine Learning: A Probabilistic Perspective, Kevin P. Murphy