AI academy - Building Multimodal AI Systems for Human-Centered Applications

This workshop bridges core AI concepts to a complete system-level, human-centered application.

The session begins with a case-study keynote that revisits sequence models (RNNs), transformer architectures, embeddings, and multimodal representation learning. These concepts are embedded within a case study focused on building an emotion-aware recommender system for paintings. Participants explore how perception and affective states can be represented computationally and how visual, contextual, and affective information can be aligned within a shared representation space.

The workshop then connects these technical foundations to human-centered applications such as museum guidance and AI-supported art therapy. Participants examine how recommendation systems can guide visual exposure and personalise experiences while preserving human oversight, professional judgement, and responsibility.

In a hands-on lab, participants interact with a simplified AI pipeline using Jupyter notebooks. They explore embeddings, compare unimodal and multimodal representations, generate personalised recommendations, and interpret model behaviour in relation to user experience.

In the afternoon, participants work in groups to design their own AI-driven systems, focusing on how learned representations can be translated into meaningful human applications. The workshop concludes with short presentations, a guided art therapy demonstration led by a professional therapist, and an immersive VR experience using personalised AI recommendations.

Morning

Keynote Lecture: From Models to Multimodal Systems

Topics covered through the keynote:

• Reframing AI from isolated models to systems interacting with humans
• Representation of human perception and emotion using Valence-Arousal space
• Embeddings and representation learning
• Sequence models (RNNs) and temporal data modelling
• Transformer architectures and attention mechanisms
• Multimodal representation learning and joint embedding spaces
• Recommender systems built on learned representations

Hands-on Lab: Multimodal Representations and Recommendation

• Explore unimodal and multimodal embeddings
• Align visual and affective representations
• Generate and inspect personalised recommendations
• Interpret model behaviour in relation to user experience
  
  

Afternoon

Group Project: User Modelling and Recommender System Design

Participants work in groups to design an AI system that uses learned representations to address a human-centered problem in healthcare, education, creativity, or cultural experience. They define:

• input data and modalities
• representation and recommendation approach
• user interaction and experience
• human oversight, risks, and limitations
  
  

Group Presentations

Short presentations (approximately 5 minutes per group)
Discussion and feedback

Experiential Session: Personalised Art and Immersive AI

Guided personalised art therapy demonstration
Interaction with AI-supported recommendations in VR
Reflection on emotional, cognitive, and experiential responses

By the end of the course, participants will be able to:

• understand how sequence models, transformers, and embeddings are used in real-world systems
• explain multimodal representation learning and its role in AI applications
• understand how recommender systems operate using learned representations
• connect AI models to human-centered applications and user experience
• critically reflect on how AI systems can shape perception and behaviour
• design conceptual multimodal AI systems for real-world use cases
• identify ethical and practical considerations in human-centered AI systems
• 
  

One-day intensive workshop combining:

• case-study keynote lecture
• hands-on coding session in Jupyter notebooks
• group-based system design exercise
• presentations and discussions
• experiential learning with art therapy and immersive VR