The application of machine learning in neuroscience has expanded markedly in recent years, propelled by advances in data acquisition technologies and the proliferation of large-scale neural, imaging, and behavioral datasets. Machine learning techniques have become indispensable for capturing complex, high-dimensional patterns inherent in brain data; however, their deployment varies significantly in the extent to which they prioritize predictive performance over mechanistic insight. This systematic review offers a comprehensive synthesis of ML-based approaches in contemporary computational neuroscience, with a particular focus on the conceptual and methodological distinction between predictive models, aimed at optimizing the decoding or forecasting of neural and behavioral outcomes, and mechanistic models, which seek to elucidate the computational principles and biological architectures underpinning brain function. A systematic review of the peer-reviewed literature was conducted, encompassing studies that applied machine learning techniques to the analysis of neural data derived from neuroimaging, electrophysiological recordings, cellular-level measurements, and multimodal experimental paradigms. The extant literature reveals a strong dominance of predictive modeling, typically utilizing supervised learning and deep neural networks to classify brain states, decode experimental conditions, or predict cognitive and behavioral phenotypes. Conversely, mechanistic machine learning models, often grounded in computational neuroscience traditions such as dynamical systems theory, probabilistic inference, and network modeling, remain comparatively underrepresented, yet they offer essential explanatory value by linking algorithmic components to biophysical mechanisms and system-level dynamics. Emerging hybrid approaches aim to reconcile predictive accuracy with mechanistic transparency, though they remain methodologically heterogeneous and are often limited by insufficient validation. This review delineates critical challenges confronting machine learning in neuroscience, including the opacity of complex models, limited reproducibility, and difficulties in cross-scale integration. It concludes by advocating the development of integrative modeling frameworks that transcend the prediction–explanation dichotomy, thereby fostering a more unified, biologically grounded understanding of brain function.