Learning with Music Signals: Technology Meets Education

音乐分发、存储和消费的革命激发了人们对开发用于组织、分析、检索和呈现音乐相关数据的技术和工具的极大兴趣。因此，音乐信息检索（MIR）领域在过去20年中已经成熟为与许多不同学科相关的独立研究领域，包括信号处理，机器学习，信息检索，音乐学和数字人文学。该项目旨在利用音乐作为一个具有挑战性和教育性的多媒体领域，在这些学科的技术和教育方面开辟新天地。该项目的独特之处在于它从不同的角度接近和探索学习的概念。首先，从数据中学习，我们将建立并推进最新的深度学习（DL）技术，直接从原始音乐信号中提取复杂特征和隐藏关系。其次，通过学习传统的工程方法的经验，我们的目标是更好地了解现有的，并通过以各种方式整合先验知识，开发更多的可解释的基于DL的系统。特别是，作为一种具有巨大潜力的新策略，我们希望将经典的基于模型的MIR方法转化为可微的多层网络，然后将其与基于DL的技术混合，形成可解释的混合模型，这些模型不易受到数据偏差和混杂因素的影响。第三，与领域专家合作，我们将考虑专门的音乐语料库，以更深入地了解音乐数据和我们的模型的行为，同时探索音乐学研究的计算模型的潜力。第四，我们将研究音乐如何作为一种激励工具，使信号处理或机器学习等技术学科的学习成为一种互动的追求。通过我们的整体学习方法，我们希望在可解释的混合模型的开发方面取得重大进展，并重塑最新技术在跨学科研究和教育中的应用和传播方式。

The revolution in music distribution, storage, and consumption has fueled tremendous interest in developing techniques and tools for organizing, analyzing, retrieving, and presenting music-related data. As a result, the field of music information retrieval (MIR) has matured over the last 20 years into an independent research area related to many different disciplines, including signal processing, machine learning, information retrieval, musicology, and the digital humanities. This project aims to break new ground in technology and education in these disciplines using music as a challenging and instructive multimedia domain. The project is unique in its way of approaching and exploring the concept of learning from different angles. First, learning from data, we will build on and advance recent deep learning (DL) techniques for extracting complex features and hidden relationships directly from raw music signals. Second, by learning from the experience of traditional engineering approaches, our objective is to understand better existing and to develop more interpretable DL-based systems by integrating prior knowledge in various ways. In particular, as a novel strategy with great potential, we want to transform classical model-based MIR approaches into differentiable multilayer networks, which can then be blended with DL-based techniques to form explainable hybrid models that are less vulnerable to data biases and confounding factors. Third, in collaboration with domain experts, we will consider specialized music corpora to gain a deeper understanding of both the music data and our models' behavior while exploring the potential of computational models for musicological research. Fourth, we will examine how music may serve as a motivating vehicle to make learning in technical disciplines such as signal processing or machine learning an interactive pursuit. Through our holistic approach to learning, we want to achieve significant advances in the development of explainable hybrid models and reshape how recent technology is applied and communicated in interdisciplinary research and education.