Computational implications of motor primitives for cortical motor learning

Motor control is a complex, high-dimensional task. It has been suggested that the brain may reduce the dimensionality of this control problem by using a set of motor primitives (or muscle synergies) that encode frequently used patterns of muscle co-activations in the time and/or weight-domain (Bizzi et al., 1991; Kutch and Valero-Cuevas, 2012; Bizzi and Cheung, 2013; Alessandro et al., 2013). In this project we investigate how motor primitives may affect the learning of movement in motor cortex.
We model motor cortex as a recurrent neural network (RNN) and compare two potential output scenarios: In the direct network, there is a direct feedforward projection from the cortical RNN to the muscle output layer (Hennequin et al., 2014; Sussillo et al., 2015). In the motor primitive network, we assume that motor cortex can control muscle activity only indirectly, via an intermediate layer of motor primitive units (MPUs). The fixed set of weight vectors from the MPUs to the muscle output layer is then a simple model of (spatial) motor primitives, encoding patterns of frequently used muscle co-activations.
For an idealized training with gradient descent, the motor primitive layer does not affect training performance. This is different for a less powerful rule, such as weight perturbation learning (Dembo and Kailath, 1990; Züge et al., 2023), which can be more easily implemented biologically: In a network trained with weight perturbation an intermediate motor primitive layer can increase training speed. This holds in linear networks with less motor primitive units than muscles and if the target muscle activations are realizable with the given motor primitive weight matrix.