Reliably identifying cell types is crucial for understanding neural systems (1). However, classifying the neurons recorded in large-scale population recordings into well-defined types remains a challenge. While some methods rely on genetic targeting or anatomy for classification, a functional approach used in sensory systems is to analyze neuronal responses to stimuli that elicit distinguishable responses across different types (2, 3, 4). Yet, previous approaches of this nature have been limited by a restricted set of stimuli and scalability issues, thus preventing comprehensive characterization of functional properties in large-scale population analysis. In the retina, recent works (4) have shown that characterizing surround suppression is necessary to match functional cell types to the ones found in transcriptomics-based and anatomical classification. However, characterizing center-surround responses requires stimuli tailored to each cell and cannot be easily scaled to population recordings. In this study, we propose a method to overcome this issue, to allow classification of large population cells recorded with multi-electrode arrays. We designed a modified version of the white noise stimulus, the Multi Spatial Frequency checkerboard (MSF), and showed that it allowed us to detect surround better than classical stimuli. We then trained spatiotemporal CNNs (5) to predict the responses of the recorded cells, and predicted the responses of individual cells to discs of different sizes, to characterize their center-surround responses. Finally, we used a classifier based on previous datasets (4) to classify these cells into different types according to their predicted responses. Our end-to-end pipeline could predict the cell types of ganglion cells with a good performance. Our study proposes an efficient method for quantifying the suppressive surround in large-scale retinal recordings, and suggests a classification strategy of cell types in large-scale recordings matched with transcriptomics and anatomy (6, 7).