Abstract
The quest to understand music’s emotional power has long fascinated philosophers, acousticians, musicians alike—not to mention the composers, performers and educators entrusted with its transmission. Ever since the development of musical notation, score-based analyses have featured prominently in exploring the role of musical structure in conveying emotion. The development of experimental psychology led to further insights, offering the ability to explore the perceptual consequences of tightly controlled musical sequences. More recently, dramatic increases in data storage and decreases in the cost of computational power offer the ability to examine large-scale changes impossible to discern from individual experiments or analyses. While each approach offers invaluable perspective, their vastly different scales poses challenges to integrating their insight. How can the question for understanding music’s emotion power best draw on these disparate tools? My team is working on a large-scale project combining analyses of scores and audio files with perceptual experiments, using advanced statistical techniques in order to “unweave” the cues for musical emotion. Grounded in widely performed sets of Preludes by Bach and Chopin, this project offers a novel opportunity to explore the evolution of musical communication across musical eras. Utilizing sets of pieces by two renowned composers, we aim to balance the power of experimental methods with the close attention to detail of score-based analyses to offer new insight into some of musical history’s most cherished compositions. We believe this approach holds exciting potential for bringing together different perspectives on musical emotion, and are looking forward to exploring ways to expand this work through greater connection with the MIR community.
Abstract
Feature attribution is often presented as the process of selecting a small subset of features most responsible for a prediction. Task-dependent by nature, precise definitions of responsibility encountered in the literature are however not always consistent. This lack of clarity stems from the fact that we usually do not have access to any notion of ground-truth attribution and from more general debates on what good interpretations are. We propose to formalise feature selection/attribution based on the concept of relaxed functional dependence. In particular, we extend our notions to the instance-wise setting and derive necessary properties for candidate selection solutions, while leaving room for task-dependence. By computing ground-truth attributions on synthetic datasets, we evaluate many state-of-the-art attribution methods and show that, even when optimised, some fail to verify the proposed properties and provide wrong solutions.
Abstract
"In this talk, first we look at how different inductive biases can be incorporated into machine learning systems, and what are the consequences of such inductive biases. We then review some of the widely-used inductive biases in deep learning models, and try to understand how they help models improve their prediction performance. More specifically, we look at popular data augmentation techniques, and their inductive biases. Further, we look at how the decision boundaries of deep neural networks can cause deep neural networks to be more vulnerable to adversarial attacks, and how this can be systematically and objectively measured. And in the end, we look at how data augmentation techniques affect decision boundary of deep neural networks, from two different aspects: 1) incorporating an inductive bias, and 2) incorporating stochasticity into training data. Using tools from statistics, we disentangle these factors, and try to analyze their effects independently, using our proposed analysis framework."
arXiv, opens an external URL in a new window
Johannes Kepler University Linz
Altenberger Straße 69
4040 Linz, Austria
Go to JKU Homepage
