Postdoc 2025-26: Lifelong learning for co-creative music generation

  • Duration: 18 months, starting March 2025.
  • Location: Lille (CRIStAL, CNRS, University of Lille); partial remote work is possible
  • Supervisor and contact: Ken DĂ©guernel (CR CNRS)
  • Applications until January 3 (open)
  • Qualifications: PhD in Computer Music, with strong experience in machine learning  and human-computer interaction.

I. Context

Recent advancements in Human-AI co-creative systems, especially within Mixed Initiative Co-Creative (MICC) frameworks, have begun to reshape musical creation, enabling new processes and interactions [1-2]. However, current systems often fall short in adapting to users over time, typically requiring unilateral adaptation — either through the user learning to operate the AI or the system being manually updated by an engineer-musician [3-4]. This internship will focus on laying the foundation for curiosity-driven learning [5] within MICC musicking [6], equipping the AI with mechanisms for adaptive, long-term engagement in creative processes [7-8].

This project sits at the intersection of several key areas, primarily Music Information Retrieval (MIR), Lifelong Learning for Long-Term Human-AI Interaction, and New Interfaces for Musical Expression (NIME). Together, these fields provide a foundation for implementing and exploring AI-driven curiosity in musical settings, ultimately facilitating expressive, intuitive engagement between musicians and adaptive AI systems.

II. Objectives

1. Implementing lifelong learning strategies to music

The objective of this postdoc is to apply state-of-the-art lifelong learning techniques for ML-based music generation models [7]. This application to music is novel and as such will pioneer the development of models, data representations, and evaluation metrics within the context of musical creation that can successfully navigate the stability–plasticity dilemma (i.e., balancing how much new input should interfere with existing knowledge). We will try different strategies based on reinforcement learning [8]. The project will then focus on agent personalisation [9] allowing the AI system to adapt based on user-specific characteristics and preferences. Areas of interest will include engagement, inclusivity, performance, and system maintenance. This research will start by focusing on single users and then be extended to allow a single agent to adapt to the needs and styles of multiple musicians simultaneously (for example, in a band setting).

This work will build upon existing machine learning models for co-composition [10] or co-improvisation [3] (but we are open to use other tools). We will first focus on developing a supervised reinforcement learning approach for music generation models based on user feedback. The methodology will involve training the music generation model through an iterative process of interaction with human users. Human users will then explicitly evaluate or provide feedback on these outputs, either online during the performance or offline as a posteriori analysis. This multi-dimensional feedback signal will be used as the reward function to fine-tune the model’s parameters through reinforcement learning algorithms. As the system improves through this interactive process, we will transition to more implicit forms of feedback based behavioural cues, and computational music analysis during the co-creative process.
The reinforcement learning framework will allow the model to continually adapt and personalise to each individual user’s unique preferences, creative voices, musical styles, and co-creative goals through the lifelong learning process. Strategies for balancing the stability-plasticity dilemma, such as elastic weight consolidation [11] or memory replay [12], will be studied to ensure the model can flexibly acquire new knowledge while preserving previously learned skills.

We will conduct an ablation study to systematically evaluate how different components and functionalities of the lifelong learning MICC system influence the overall co-creative experience [13]. Key stages in the music generation process such as elaboration (expanding on musical ideas) and reflection (evaluating and providing feedback) will be identified and selectively ablated or removed. This controlled analysis will isolate the impact of each component, allowing researchers to pinpoint elements that are crucial for enabling specific creative outcomes.

2. Evaluation with an ablation study

The ablation study will investigate how different system capabilities affect core aspects of co-creativity, including long-term vs short-term creative focus, depth of elaboration vs breadth of exploration, and locus of agency (human-initiated vs AI-initiated creative actions). For example, removing reflection capabilities may hinder long-term creative trajectories, while ablating elaboration functions could stifle local improvisational creativity.
By systematically removing and restoring different system components across multiple creative sessions, this analysis will map out the relationships between technological affordances and creative experiences. This will not only allow for targeted improvements to the MICC system based on desired creative objectives, but also provide a deeper understanding of how humans and AI can optimally collaborate through complementary roles facilitated by the right technological capabilities. Findings will shed light on design principles for lifelong co-creative AI systems.

The evaluation process will involve a diverse group of expert musicians, who will interact with the AI systems over an extended period. The musicians will be asked to provide regular feedback on their experiences, focusing on aspects such as the system’s adaptability, responsiveness, and ability to generate musically coherent and engaging output.

References

[1] Jordanous (2017). Co-creativity and perceptions of computational agents in co-creativity. International Conference on Computational Creativity.

[2] Herremans et al. (2017). A functional taxonomy of music generation systems. ACM Computing Surveys, 50(5).

[3] Nika et al. (2017). DYCI2 agents: merging the ‘free’, ‘reactive’ and ‘scenario-based’ music generation paradigms. International Computer Music Conference.

[4] Lewis (2021). Co-creation: Early steps and future prospects. Artisticiel/Cyber-Improvisations.

[5] Colas et al. (2020). Language as a cognitive tool to imagine goals in curiosity driven exploration. NeurIPS.

[6] Small, C. (1998). Musicking: The meanings of performing and listening. Wesleyan University Press.

[7] Parisi et al. (2019). Continual lifelong learning with neural networks: A review. Neural networks, 113.

[8] Scurto et al. (2021). Designing deep reinforcement learning for human parameter exploration. ACM Transactions on Computer-Human Interaction, 28(1).

[9] Irfan et al. (2022). Lifelong Learning and Personalization in Long-Term Human-Robot Interaction (LEAP-HRI). *ACM/IEEE International Conference on Human-Robot Interaction.

[10] Sturm et al. (2016). Music transcription modelling and composition using deep learning. arXiv:1604.08723.

[11] Kirkpatrick et al. (2017). Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13).

[12] Ho et al. (2023). Prototype-guided memory replay for continual learning. IEEE Transactions on Neural Networks and Learning Systems.

[13] Lin et al. (2023). Beyond prompts: Exploring the design space of Mixed-Initiative Co-Creativity Systems. International Conference on Computational Creativity.