| dc.description.abstract | The quest to understand biological intelligence and to develop artificial intelligence (AI) is a long-standing academic pursuit. While past work in AI drew inspiration from neuroscience, AI models can also be used to help understand brain function. One of the fundamental open questions in neuroscience is understanding how the central nervous system enables the learning and control of body movements. In this thesis, we used AI models to understand how a network of neurons can instantiate sensorimotor computations. In the first part of the thesis, we developed embodied AI agents that learn to control a biomechanically realistic 3D virtual human arm model with 50 muscles simulated under the influence of gravity. Specifically, we incorporated ideas from computational theories of human motor control, developmental psychology (body babbling), and robotics to construct four types of internal models for different sensorimotor transformations (i.e., forward kinematics, inverse kinematics, forward dynamics, and inverse dynamics) using artificial neural networks (ANNs) to produce goal-directed movements.
In the second part of the thesis, we performed an in-silico characterization of the inverse dynamics (ID) neural network and systematically compared its internal representations and dynamics to those observed in the primary motor cortex (M1). Our analyses revealed several key parallels: directional tuning of units with preferred directions (PDs) for movement during 2D and 3D center-out reaching tasks, postural dependence of PDs, non-uniform PD distributions, dynamic shifts in PDs over time, encoding of movement direction in population activity trajectories in low-dimensional latent space, and the emergence of rotational dynamics, all phenomena previously reported in M1. Furthermore, the ID network analyses also offered some mechanistic insights and predictions. First, the average peak muscle excitation (APME) predicted the dominant axis of the PD distribution, suggesting that the non-uniform PD distribution arises from anisotropies in the limb's biomechanical properties. Second, this non-uniform PD distribution was reflected in the latent space population trajectories, which clustered into two distinct groups corresponding to movement directions near each mode of the PD distribution. Third, we demonstrated that rotational dynamics in the ID network, and possibly M1, can emerge not only from intrinsic recurrent connectivity but also from external inputs such as task goals and sensory feedback.
Finally, we also assessed the causal relevance of directional tuning by performing single-unit and population lesion experiments targeting directionally selective units in the ID network. These lesions induced directionally specific changes in motor behavior and latent space trajectories, underscoring the causal role of these units in generating behavior and shaping population dynamics. Additionally, we found that for small populations, the behavioral and latent space effects of population lesions could be approximated by the linear summation of single-unit lesion effects. However, this approximation may not hold for larger populations.
Overall, this thesis leverages the emerging synergy between neuroscience, biomechanics, developmental psychology, robotics, and AI to provide insights into the biological basis of movement control. This work also has practical implications for designing intelligent robotic systems, brain-machine interfaces (BMIs), and the development of neural co-processors for rehabilitation. | en_US |
