TL;DR
This paper introduces BAND, a novel semi-supervised model that effectively captures subtle, transient neural signals responsible for motor corrections during reaching. By combining latent dynamics with behavioral supervision, BAND outperforms existing methods in predicting both planned movements and unexpected adjustments, revealing that these corrections are encoded in small, feedback-driven neural variability.
What Problem Does This Paper Address?
Understanding how the brain controls movement is a fundamental challenge. While we can model the "planned" part of a movement – the preparation and execution of a learned trajectory – it's much harder to account for the rapid, online adjustments the brain makes when unexpected things happen, like encountering a force field. These corrections are crucial for smooth and adaptive behavior. Previous methods, often relying on unsupervised learning to find patterns in neural activity, tend to overlook these corrections because they manifest as small, transient deviations in neural firing that can be mistaken for noise. This paper aims to bridge this gap by developing a model that can reliably detect and interpret these subtle, yet behaviorally significant, neural signals.
Methods
The core of this work is the Behavior-Aligned Neural Dynamics (BAND) model. BAND builds upon the established LFADS (Latent Factor Analysis via Dynamical Systems) framework, which uses sequential autoencoders to infer latent dynamics from neural activity. The key innovation in BAND is the addition of behavioral supervision. Instead of solely focusing on reconstructing neural activity (as in LFADS), BAND is trained to simultaneously predict both neural activity and the actual observed behavior. This is achieved through a semi-supervised learning approach, where a sequence-to-sequence behavior decoder is integrated into the model. This allows BAND to learn latent representations that are not only consistent with neural firing patterns but also strongly predictive of behavior, even when those neural signals are small and transient. The authors also explore a linear dynamical model with behavior supervision (PSID) and a purely supervised method (CEBRA) for comparison.
Datasets
The study utilizes two main types of datasets:
- Neural Latents Benchmark (NLB) hand-reaching datasets: These datasets, recorded from monkeys performing various hand-reaching tasks, are used to evaluate the general performance of BAND against other models in predicting behavior and neural activity.
- Primate motor cortex recordings during center-out reaches with a force-field perturbation: This dataset, specifically designed to elicit and study motor corrections, is central to the paper's findings. It includes trials with baseline movements, adaptation to a force field, and washout periods.
Key Insights and Ideas
- Small neural variability encodes large behavioral changes: The most striking finding is that motor corrections, which lead to significant behavioral adjustments, are encoded in neural activity that is surprisingly small and transient. Unsupervised methods, focused on explaining the bulk of neural variance, tend to discard these subtle signals.
- Behavioral supervision is key for capturing corrections: Simply modeling neural dynamics isn't enough. Adding a behavioral supervision signal to the model is crucial for forcing it to attend to and learn the neural patterns that drive these corrections.
- Motor corrections are feedback-driven: The research strongly suggests that these online corrections are not pre-programmed but are triggered by sensory feedback about the movement's outcome. This feedback loop appears to involve a delay of approximately 90 ms before influencing neural activity.
- M1, not PMd, is critical for corrections: While both primary motor cortex (M1) and dorsal premotor cortex (PMd) are involved in movement, the study indicates that M1 neurons are specifically involved in encoding these feedback-driven corrections, while PMd is more associated with the overall movement plan.
- A revised view of motor control: The findings challenge a purely feedforward view of motor control, emphasizing the dynamic interplay between planned movements and continuous, feedback-driven adjustments.
Key Results
- BAND achieves superior behavior prediction: BAND significantly outperforms other non-ensemble models on the NLB benchmark for hand velocity reconstruction, achieving R² of 92% in the force-field perturbation task, matching the performance of a powerful bi-directional RNN oracle.
- LFADS struggles with corrections: The unsupervised LFADS model, while good at capturing planned movements, could only predict the presence of 5 Hz oscillations in hand velocity but not their phase, indicating a failure to capture the timing of corrections.
- BAND captures 5 Hz oscillations and their phase: BAND successfully reconstructs both the amplitude and phase of the characteristic 5 Hz oscillations observed during adaptation to the force field.
- Controller is essential for corrections: Ablating the controller component of BAND (which is trained with behavioral supervision) drastically reduces behavior prediction accuracy (from R²=91% to R²=71%), highlighting its importance for capturing corrections.
- Feedback loop identified: Analysis of BAND's decoder weights reveals a consistent lag of approximately 90 ms between neural activity and behavioral feedback, suggesting a sensory feedback loop influencing motor cortex activity.
Limitations and Future Work
- Limited to trained animals and specific tasks: The current study focuses on highly trained animals performing well-defined reaching tasks. Future work could extend BAND to study motor learning and error processing in less trained animals or more naturalistic, unstructured behaviors.
- Real-time applications: While the current analysis is on structured, offline data, the authors suggest that BAND's approach could be extended to real-time applications, such as brain-computer interfaces.
- Generalizability across animals: The paper mentions future work to extend BAND to identify shared latent dynamics across multiple animals, which would be valuable for studying generalizable neural representations.
- Investigating plasticity: The authors note that understanding the mechanisms of adaptation to force-field perturbations, which may differ from visuomotor perturbations, requires further investigation.
Summary updated