KAD: No More FAD! An Effective and Efficient Evaluation Metric for Audio Generation

📎https://arxiv.org/pdf/2502.15602

#audiogeneration, #audioqualityevaluation, #FréchetAudioSistance, #maximummeandiscrepancy, #kernelMethods

Consider a Transformer as a kernel machine. Look at the extension of kernels to an asymmetric form. Motivated by empirical observation of low rank properties of self-attention, find a link between kernel Singular Value Decomposition and self attention.

#MachineLearning #KernelMethods #Attention #Transformers

https://arxiv.org/abs/2305.19798