Recent advances in recording techniques and behavioral monitoring are revolutionizing systems neuroscience, enabling the study of behavior in naturalistic settings. Using fiber photometry data from dopamine axons (DA) in the striatum, we set out to characterize the homogeneity and heterogeneity across striatal areas during naturalistic behavior. These recordings were performed while a mouse explored multiple items placed inside a box, with which they could interact freely. Specifically, our goal was to test the hypothesis that DA axons in the ventral striatum (VS) solely encode value information, while those in the dorsal striatum (DS) multiplex it with other outcome and region-specific events, such as movement and threat. However, testing this hypothesis in unstructured naturalistic experiments is challenging with existing state-of-the-art tools. Consequently, we used a newly developed interpretable deep learning framework called DUNL suited to settings that lack trial structure and data-limited regimes. Based on convolutional sparse autoencoders (SAEs), DUNL deconvolves multiplexed neural signals into temporally localized and sparse components. SAEs have become a popular method for interpreting LLMs; here we demonstrate their potential for understanding neural computation. The encoder is based on a deconvolutional residual network with ReLUs that identifies sparsely localized events. The decoder is a convolutional network that models the neural data as the linear combination of learned localized kernels and the sparse code from the encoder. We quantified reconstruction errors as a function of the number of kernels and sparsity level, a common method for model selection in SAEs. The most stable model fit for each optical fiber indeed suggests that VS is characterized by one kernel, while DS is best characterized by 2 different kernels. Our analysis highlights the unique capabilities of convolutional SEAs to uncover the neural mechanisms underlying behavior under naturalistic conditions in unstructured, data-limited experiments.