I dunno why I find the idea of "finding the hyperplane that maximally separates classes" so unpleasant compared to the approach that makes sense to me "optimize the weights of a parameterized function to minimize an objective function". I imagine under the hood they're not really different. Generally, I think SVMs appear to statisticians while deep learning appeals to physicists and probabilicians.