Mapping and understanding the functional differences between neuron types is crucial for advancing our knowledge of neural circuits and brain function. However, classifying neuron types in vivo remains challenging due to the difficulty of collecting reliable labels and the complexity of deciphering how neuronal activity varies across different conditions and stimuli. Even when labeled data is available, it is often hard to identify subtle functional differences between neuron classes. To address these challenges, we introduce a novel spatiotemporal Transformer-based architecture that learns a soft clustering of neural responses both over time and across trials. This grouping structure, captured by our group token Transformer, allows the model to discover latent patterns in neuronal activity that reveal functional differences across cell types. We applied this model to calcium imaging datasets from the olfactory cortex and successfully classified distinct projection neuron classes---those projecting to the olfactory bulb (feedback neurons) and those projecting to the medial prefrontal cortex (feedforward neurons). By uncovering differences in the timing and coordination of neuronal firing across these projection classes, which are difficult to detect with conventional methods, our model provides a powerful new tool for functional cell type classification. This approach not only enhances our ability to map neural circuits but also holds potential for broader applications in other model systems, offering deeper insights into how neuronal populations function and interact across different brain regions.