InFoRM: a unified inverse and forward model for sensorimotor control.

Area of Science:

• Computational Neuroscience
• Robotics
• Motor Control

Background:

• Traditional sensorimotor control models utilize separate inverse and forward internal models.
• The neural basis for separating these models remains unclear, and separate networks may increase computational cost.
• Investigating unified models could offer computational advantages and insights into neural processing.

Purpose of the Study:

• To investigate if inverse and forward sensorimotor functions can be integrated within a single neural circuit.
• To introduce and evaluate an inverse-forward-recognition model (InFoRM) for sensorimotor control.
• To compare the performance and resource efficiency of InFoRM against classical separated models.

Main Methods:

• Implemented InFoRM using neural networks.
• Compared InFoRM with control architectures based on separated inverse and forward models.
• Utilized recorded 3D kinematics for desired trajectories and inverse dynamics to derive efferent and afferent signals.

Main Results:

• InFoRM significantly outperformed control architectures in reproducing cyclic reaching movements across various conditions.
• The InFoRM required fewer computational resources compared to separated models.
• InFoRM demonstrated generalization capabilities, morphing to untrained movement directions and generating novel motor commands and predicted feedback.

Conclusions:

• Integrating inverse and forward sensorimotor processes within a single neural network (InFoRM) offers significant computational advantages.
• Unified sensorimotor models may be more efficient and flexible than traditional separated models.
• These findings suggest that exploring unified neural circuits for sensorimotor control is a promising research direction.