Journal: bioRxiv

Article Title: Task-driven neural network models predict neural dynamics of proprioception

doi: 10.1101/2023.06.15.545147

Figure Lengend Snippet: A. We created a normative framework to study the neural code of proprioception. Using synthetic muscle spindle inputs derived from musculoskeletal modeling, we optimized deep neural networks to solve different computational tasks in order to test hypotheses (N=16 tasks). Task-driven models give rise to learned representations across the artificial neural network models. B. We interrogated which type of hypothesis creates models that best generalize to explain the activity of neurons in the cuneate nucleus (CN) of the brainstem and in the primary somatosensory cortex (S1, area 2) of non-human primates performing a center-out reaching task comprising active and passive movements.

Article Snippet: Using these end-effector trajectories, we generated realistic proprioceptive stimuli as done for the data from the center-out reaching task using the same OpenSim musculoskeletal model of the macaca mulatta upper limb ( ).

Techniques: Derivative Assay, Activity Assay

Journal: bioRxiv

Article Title: Task-driven neural network models predict neural dynamics of proprioception

doi: 10.1101/2023.06.15.545147

Figure Lengend Snippet: A. Diagram of the computational pipeline to compute realistic synthetic proprioceptive inputs from 3D character trajectories. In brief, 2D character trajectories are augmented and projected into 3D space using a 2-link 4 degree of freedom (DoF) arm model by randomly selecting candidate starting points in the workspace of the NHP arm. From those movements, muscle lengths and velocities are computed using inverse kinematics and musculoskeletal modeling. B. Left: Distribution of joint angles and trajectories in the behavioral (top) and synthetic (bottom) data, illustrating that the movement statistics in the synthetic dataset are designed to (broadly) encompass the biological movements of NHPs during the center-out reaching task. Right: workspace trajectory starting points of the synthetic dataset (blue), encompassing the behavioral workspace (red). Note that experimental centerout trajectories themselves are not part of the synthetic dataset. C. In total, 16 computational tasks were designed to reflect different hypotheses about functional proprioceptive processing. Each hypothesis contains one or several objectives, grouped by similarity. The background of the panel is color coded based on the hypothesis and it will be used throughout the manuscript. For each task, the learning objective is highlighted in red with each arm pictogram. D. Test performance of each network model, on selected tasks (N = 350 models except the autoencoder task where N=295), with respect to the number of layers (i.e. model depth). For the regression tasks, we used the mean squared error (MSE), for the action recognition task, the classification accuracy, for the autoencoder task, the relative error for the muscle length and for the redundancy reduction task, the Barlow loss (See methods). Four types of deep neural network architectures were designed to integrate proprioceptive signals in different ways: spatial-temporal, temporal-spatial, and spatiotemporal TCNs, and spatial-LSTM. The color code of each point reflects the architecture type.

Article Snippet: Using these end-effector trajectories, we generated realistic proprioceptive stimuli as done for the data from the center-out reaching task using the same OpenSim musculoskeletal model of the macaca mulatta upper limb ( ).

Techniques: Functional Assay

Journal: bioRxiv

Article Title: Task-driven neural network models predict neural dynamics of proprioception

doi: 10.1101/2023.06.15.545147

Figure Lengend Snippet: A. Correlation between test and train neural explainability (NHP S; CN) with respect to the number of PCs for representative models trained on the action recognition task (N=6 models, two per TCN subtype). Shaded are is the confidence interval at 95%. We note that we used 75 principal components throughout the manuscript (for fair comparison with the number of proprioceptive inputs) as it avoids overfitting but higher test EV is possible with around 100 dimensions. B. Neural predictions from TCNs workflow. Top: task optimization of neural network models is done using the large-scale synthetic proprioceptive input data (derived from pose estimation and musculoskeletal modeling). Bottom: experimental proprioceptive inputs during center-out reaching of NHPs is used as new inputs to the frozen neural network models to generate model activation per layer, which are combined using principal component analysis. This constitutes model features, which we generate for both untrained, random models and task-trained models. From these features, we generate neural predictions using spatial linear regression (i.e., weights do not change for each time bin). C. Distributions of explained variance scores between training and testing trial datasets for each NHP S (CN, top row) and NHP H (S1, bottom row) obtained from models trained on the hand position and velocity task for the active condition (left column) and passive one (right column). For visualization, N = 20000 randomly sampled single-neuron predictions are shown across models architectures, model layers and neurons. D. Pairwise comparison of single-neuron explained variance between layers of one example deep neural network model (same as in ) and baseline linear encoding model using muscle spindle inputs for NHP S (CN, left column; N = 47 neurons) and NHP H (S1, right column; N = 27 neurons) during active movements. Each row represents the comparison performed for the specific layer of the network. Statistical significance was computed with Wilcoxon signed-rank test (*** = p < 0.0001; ** = p < 0.001; * = p < 0.05). Deep layers of task-driven models are the ones that significantly outperform linear models. E. Same as in panel D but for passive trials.

Article Snippet: Using these end-effector trajectories, we generated realistic proprioceptive stimuli as done for the data from the center-out reaching task using the same OpenSim musculoskeletal model of the macaca mulatta upper limb ( ).

Techniques: Comparison, Derivative Assay, Activation Assay