Content area
Recent advances in neurotechnology have enabled the simultaneous recording of large-scale neural activity and behavioral data, opening opportunities to elucidate the neural mechanisms underlying behavior and to improve brain-computer interfaces (BCIs). Establishing an association between neural and behavioral recordings plays a pivotal role in accomplishing these opportunities. However, both neural and behavioral recordings are inherently high-dimensional, posing challenges to directly establish the association. Alternatively, projecting these high-dimensional recordings onto structured latent spaces can be effective in identifying underlying neural and behavioral relationships and bridging the gap between the two.
Algorithms that project neural and behavioral recordings onto lower-dimensional representations have been proposed. For instance, latent variable models (LVMs) have been introduced to transform neural data into interpretable low-dimensional representations for analysis. While effective, these methods often overlook the sequential and causal properties inherent in neural activity. Meanwhile, approaches for automatic behavior understanding typically depend on extensive human annotation to achieve high precision. The annotation procedure is often labor-intensive and subjective, limiting their scalability and consistency. To overcome these limitations, this thesis proposes approaches that enable a more biologically realistic and efficient analysis of neural and behavioral recordings.
Specifically, we introduce a neural representation learning approach explicitly incorporating temporal causality. In this framework, representation learning is formulated to estimate future neural activity solely from its past. We additionally included a graphical prior to model pairwise spatial interactions among neural recording channels. Experiments on synthetic and actual neural datasets demonstrate that this method can enhance the estimation of future neural dynamics, recover the underlying spatial interactions, and align neural trajectories with behavioral states. In parallel, for efficient behavioral state discovery, we developed an active learning-based, semi-supervised approach for behavioral state discovery and classification, significantly reducing the annotation burden while maintaining high accuracy. As a result, we introduce OpenLabCluster, an open-source toolkit designed to identify behavior categories across diverse species to democratize these behavioral techniques as a convenient user interface. Finally, we propose algorithms to bridge neural representations and behavioral states. With an application to brain-to-text decoding, we show enhanced accuracy in decoding neural signals into phonemic and textual outputs by leveraging refined behavioral states. This algorithm underscores the importance of using relevant and precise behavioral states as additional guidance to enhance brain decoding fidelity.
Overall, these contributions facilitate the understanding of neural-behavior relationships and improve the precision of brain-computer interfaces, potentially paving the way for neural computation discovery and the development of high-performance BCIs.