Deep Kernel Machines

点击翻译按钮获取中文摘要

Deep Kernel Machines (DKMs) represent a novel category of kernel methods that share similarities with neural networks (NNs) in their ability to learn effective data representations. However, unlike NNs, DKMs acquire these representations exclusively through kernels, bypassing inscrutable weight matrices. This unique characteristic offers an opportunity to gain insights into the mechanisms underlying deep learning models. Recent studies have demonstrated that convolutional DKMs can achieve competitive performance, comparable to models like ResNets, on image classification tasks such as CIFAR-10. This suggests the potential of DKMs as a practical alternative to neural networks. This research project aims to broaden the scope of DKMs by applying them to diverse dataset domains, including graphs through the Graph Deep Kernel Machine and sequences via Deep Kernel Transformers. The objectives of this research are to refine DKMs into a practical methodology and to illuminate novel perspectives on the workings of deep models through the lens of kernel-based learning. This project falls within the EPSRC artificial intelligence and robotics/mathematical sciences research area