We conceptualize and implement a novel framework for classifying seismograms using the FastMap algorithm (Faloutsos & Lin, 1995) and Support-Vector Machines (SVMs). Most machine-learning applications for detecting earthquakes to date have used Convolutional Neural Networks (CNNs). Here we present an alternative approach. Model training comprises two basic steps: (a) Using the FastMap algorithm, we embed a database of previously classified seismograms into N-dimensional Euclidean space in a way that optimally preserves the distance between seismograms, as quantified by an arbitrary distance function measuring the dissimilarity between seismograms—this step is most easily understood as feature extraction. Then, (b) we train an SVM using these embedded seismograms, classified as either recording or not recording an earthquake. Once the SVM has been trained, a test seismogram with unknown class can be embedded in the same N-dimensional Euclidean space, and its class can be predicted by the SVM. The proposed framework has at least three conceptual advantages over CNNs: (a) domain expertise can be incorporated into feature extraction via the distance function, (b) representative elements of the training data are retained explicitly (without abstraction) and subsequently used to extract features from test seismograms, and (c) it maintains computational efficiency by avoiding expensive convolution operations. At the conference, we will present test results obtained using various distance functions and SVM kernels.