端到端语音模型学到了哪些关于说话人、语言和信道的信息？一项逐层与神经元级分析
What do End-to-End Speech Models Learn about Speaker, Language and Channel Information? A Layer-wise and Neuron-level Analysis
AI总结：
本文通过逐层与神经元级探测分析端到端预训练语音模型，揭示说话人、语言和信道信息的分布、冗余性、层级定位及性别偏差，并发现CNN在编码多种属性上与Transformer相当。
摘要：
深度神经网络本质上是不透明的，难以解释。与基于手工特征的传统模型不同，我们很难理解这些模型内部学到了哪些概念以及这些概念如何相互作用。这种理解不仅对调试至关重要，而且对确保伦理决策中的公平性也至关重要。在本研究中，我们使用探测框架[1]对预训练语音模型进行了事后功能可解释性分析。具体而言，我们分析了为说话人识别和方言识别等多种任务训练的语音模型的话语级表示。我们进行了逐层和神经元级分析，探测说话人、语言和信道属性。本研究旨在回答以下问题：i）表示中捕获了哪些信息？ii）这些信息是如何被表示和分布的？iii）我们能否识别出包含这些信息的最小子网络？我们的结果揭示了若干新发现，包括：i）信道和性别信息分布在整个网络中；ii）就某一任务而言，这些信息在神经元中是冗余可用的；iii）复杂属性（如方言信息）仅在面向任务的预训练网络中被编码；iv）并且这些信息集中在上层；v）我们可以提取出编码预定义属性的最小神经元子集；vi）显著神经元有时在不同属性之间共享；vii）我们的分析揭示了网络中存在的偏差（例如性别偏差）。我们的跨架构比较表明：i）预训练模型捕获了说话人不变性信息；ii）CNN模型在编码多种尚未充分研究的属性方面与Transformer模型具有竞争力。
英文摘要：
Deep neural networks are inherently opaque and challenging to interpret. Unlike hand-crafted feature-based models, we struggle to comprehend the concepts learned and how they interact within these models. This understanding is crucial not only for debugging purposes but also for ensuring fairness in ethical decision-making. In our study, we conduct a post-hoc functional interpretability analysis of pretrained speech models using the probing framework [1]. Specifically, we analyze utterance-level representations of speech models trained for various tasks such as speaker recognition and dialect identification. We conduct layer and neuron-wise analyses, probing for speaker, language, and channel properties. Our study aims to answer the following questions: i) what information is captured within the representations? ii) how is it represented and distributed? and iii) can we identify a minimal subset of the network that possesses this information? Our results reveal several novel findings, including: i) channel and gender information are distributed across the network, ii) the information is redundantly available in neurons with respect to a task, iii) complex properties such as dialectal information are encoded only in the task-oriented pretrained network, iv) and is localised in the upper layers, v) we can extract a minimal subset of neurons encoding the pre-defined property, vi) salient neurons are sometimes shared between properties, vii) our analysis highlights the presence of biases (for example gender) in the network. Our cross-architectural comparison indicates that: i) the pretrained models capture speaker-invariant information, and ii) CNN models are competitive with Transformer models in encoding various understudied properties.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
