Machine Learning Methods for Studying Latent Neural Activity Dynamics

This survey comprehensively maps the evolution of machine learning methods for decoding neural activity, from classical state-space models to modern deep generative approaches. It organizes techniques across three domains—single-region dynamics, multi-region communication, and behavior-aligned modeling—while highlighting emerging foundation models and open challenges in causal inference for brain research.

This academic survey documents a significant maturation in neuroscience-AI convergence, where machine learning has become essential infrastructure for interpreting brain function. The progression from linear dynamical systems to transformer-based neural foundation models reflects broader trends in AI: increasing model complexity, shift toward self-supervised pre-training, and growing emphasis on interpretability. The paper's organization across three domains reveals how researchers systematically tackle different facets of neural complexity, from local circuit dynamics to distributed communication patterns.

The relevance extends beyond academic neuroscience. Decoding latent neural dynamics has direct applications in brain-computer interfaces, neurodegenerative disease monitoring, and understanding consciousness mechanisms. As brain recording technologies improve—evidenced by recent advances from companies like Neuralink and academic institutions—demand for robust decoding tools intensifies. The survey's emphasis on behavior-aligned modeling and contrastive learning indicates industry movement toward actionable neural insights rather than pure descriptive models.

A critical gap the survey identifies is causal inference: determining directionality of communication between brain regions. This remains a bottleneck for translating decoded dynamics into mechanistic understanding. The inclusion of large-scale foundation models suggests the field is adopting transfer learning approaches similar to NLP and computer vision, potentially accelerating research velocity. However, benchmarking and evaluation standardization remain underdeveloped, limiting reproducibility and cross-study comparison. Future progress likely depends on whether researchers can reconcile model complexity with interpretability—a tension that currently favors predictive performance over explanatory power in many applications.

• →Machine learning methods for neural decoding have evolved from linear models to deep generative and foundation models, enabling analysis of large-scale brain populations.
• →The field organizes around three complementary approaches: single-region dynamics, multi-region communication patterns, and behavior-aligned task modeling.
• →Neural foundation models using transformer and diffusion architectures represent an emerging frontier that requires large-scale pre-training for cross-subject generalization.
• →Causal inference and directional communication mapping remain major open challenges limiting mechanistic understanding of brain dynamics.
• →Standardized benchmarks and evaluation criteria are underdeveloped, creating reproducibility gaps in the neuroAI research ecosystem.