Machine Learning Methods for Studying Latent Neural Activity Dynamics

A new survey paper on arXiv catalogs the machine learning techniques used to decode latent neural activity, organizing the field into three domains: single-region dynamics, multi-region communication, and behavior-aligned modeling [1][2]. The paper, submitted on June 9, 2026, traces the evolution of Latent Variable Models from early state-space approaches to modern deep generative models [1][2]. The authors structure their review around three interconnected research areas. The first, Single-Region Latent Dynamics, covers models ranging from linear dynamical systems to more complex representations using Recurrent Neural Networks and Neural Ordinary Differential Equations [1][2]. RNNs, which are designed to process sequential data, have been a foundational architecture for modeling time-series information such as neural signals [4]. The second domain, Multi-Region Communication, examines probabilistic and subspace methods for tracking information flow across brain areas, accounting for factors such as synaptic propagation delays and network connectivity [1][2]. The third, Behavior-Aligned Modeling, uses supervised or contrastive learning to separate neural activity tied to task performance from other internal states [1][2]. The survey also addresses large-scale neural foundation models, including Transformers and diffusion models, which depend on large-scale pre-training to perform well across different subjects [1][2]. Transformer architectures, which introduced attention mechanisms for modeling long-range dependencies, now underpin large language models, while diffusion models have become central to image generation systems [4][6]. The paper concludes by identifying open challenges, notably the need to establish causal links and determine the directionality of communication between brain regions [1][2]. The work appears on arXiv, the open-access e-print repository that has hosted scientific papers since 1991 and now receives roughly 24,000 submissions per month [10]. The survey is accompanied by experimental community tools under the arXivLabs framework, which provides third-party features such as citation explorers and code finders while enforcing user data privacy [8][9].