Cool project that I worked on during Fall 2025 for CMSC838B: Differentiable Programming under Dr. Ming Lin @ UMD! Access the full report here: https://www.overleaf.com/read/tkgcrfxptrzj#61e195
The PDF report is also included in this repo!
In this work, we present a novel method for generating personalized violin finger- ings for symbolic sheet music, given user preferences. Recent work has explored this problem via statistical methods or BLSTM networks, but none have explored the powerful transformer architecture — renowned for excelling at processing sequential data. In this work we discuss a method for converting symbolic sheet music (in MusicXML format) into a differentiable format compatible with trans- former architecture. Then, we explore and iteratively refine different transformer architectures to generate personalized fingerings, utilize LoRA + prefix tuning for finetuning on preferences, and present qualitative results of our method work- ing reasonably well in practice. However, we find that biggest limitation for any method in solving this problem is the lack of a comprehensive violin fingering dataset, so future directions should develop a pipeline to obtain that data, or use self-supervision to take advantage of unlabeled data (as some related works do).