Hybrid Losses for Hierarchical Embedding Learning
Haokun Tian 1 Stefan Lattner 2 Brian McFee 3 Charalampos Saitis 1 · 2025
Read Hybrid Losses for Hierarchical Embedding Learning by Haokun Tian 1 Stefan Lattner 2 Brian McFee 3 Charalampos Saitis 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-12796.txt.
Description
Haokun Tian, Stefan Lattner, Brian McFee, and Charalampos Saitis improve how learning systems sort sounds into fine-grained classes. Their paper appeared in 2025. Standard training treats every wrong answer as equally wrong, even when the wrong class is a close cousin of the right one. The team uses the tree of instrument families to teach closeness. They blend two training goals, so the model's inner map places similar sounds near each other.
They also propose measures for how well that inner map is structured and whether it can place sounds from classes never seen in training. Tests used a data set of instrument sounds with nearly 200 detailed categories in four levels. The blended goals beat earlier work on sorting, search, map structure, and unseen classes. The work matters for music tools that must handle rare instruments.
Catalog record
| CREATOR | Haokun Tian 1 Stefan Lattner 2 Brian McFee 3 Charalampos Saitis 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 22 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-12796