As high-energy nuclear physics research enters a phase characterized by multi-dimensional and highly complex data analysis, deep learning techniques are gradually becoming essential tools for understanding nuclear matter behavior under extreme conditions. This shift is driving a fundamental transformation in research paradigms from experience-driven approaches toward data-driven methodologies. This article briefly reviews the evolution of machine learning in this field, emphasizing recent advancements involving deep learning techniques. Early research (from the late 20th century to the 2010s) primarily employed traditional algorithms such as artificial neural networks and support vector machines. These studies validated the feasibility of machine learning approaches in nuclear physics through tasks like nuclear mass prediction and phase transition identification. However, due to limitations in manual feature extraction and computational capabilities, such methods did not yet extend to autonomous exploration of physical features. In the deep learning era (2010s to present), researchers have innovatively introduced point-cloud neural network architectures, enabling direct processing of final-state particle four-momentum data. This advancement has overcome the constraints of traditional methods that relied heavily on manually constructed statistical observables and initiated a conceptual leap from superficial data representations toward intrinsic physical insights. Simultaneously, unsupervised learning methods have shifted research focus from hypothesis validation to autonomous, data-driven discovery of physical laws, facilitating not only sensitive detection of anomalous signals but also opening new avenues for investigating emergent physical phenomena. Looking ahead, from developing deep learning algorithms incorporating physical priors to enhance the model physical interpretation, to meta-learning and self-supervised frameworks deepening rare event analysis; from quantum machine learning accelerating high-dimensional feature extraction, to generative models reconstructing the physical data ecosystem, these advancements will potentially propel high-energy nuclear physics research from the passive interpretation of observational data toward active discovery of physical laws, shifting analysis from fragmented, local feature exploration toward holistic comprehension of systemic behaviors. Ultimately, this progression may pave the way toward constructing an intelligent physics research system capable of autonomous knowledge discovery.