Bio
Teresa Pelinski is a PhD student at the Centre for Doctoral Training in AI and Music, part of the Centre for Digital Music at Queen Mary University of London, where she is also part of the Augmented Instruments Lab. Teresa holds a BSc in Physics from Universidad Autónoma de Madrid and a MSc in Sound and Music Computing from Universitat Pompeu Fabra (Barcelona). Her PhD is supervised by Prof Andrew McPherson (Imperial College London) and Prof Rebecca Fiebrink (University of the Arts London) and supported by UKRI and Bela.
Research interests
Digital musical instruments are commonly built using embedded hardware platforms (low-powered single-purpose microcomputers) due to their low latencies, small size and affordable prices. Many of these platforms also offer an entry point for beginners into physical audio computing through APIs that facilitate interfacing with sensors and actuators and real-time interaction. However, deploying neural networks on embedded devices is an arduous task: often, there exist no platform-specific instructions, and the limited computational resources available on-device result in large compilation times and complicate real-time inference.
Teresa's research focuses on tackling these issues to ease prototyping and experimentation with neural networks in embedded hardware in the context of digital musical instrument design. With an emphasis on reproducible workflows, she is developing tools for lowering the entry barriers of embedded neural network deployment for makers and creatives. Given the limited computational capabilities of embedded systems and hence, the need for lighter neural networks, Teresa is particularly interested in the balance between their accuracy and utility (or efficiency). The mainstream trend in deep learning is large models that generalise well with high accuracy metrics, yet in other domains, such as creative practice, the accuracy or optimality of a network might be less desirable than its utility or efficiency. Moreover, lighter neural networks are relevant not only for embedded applications but also for their lower energy consumption.