P029 Spiking Neural Networks for Controlling a Biomechanically Realistic Arm Model
Philip Bröhl*1,2, Junji Ito2, Ira Assent1,3, Sonja Grün2,4,5
1Institute for Advanced Simulation (IAS-8), Research Center Juelich, 52425 Jülich, Germany 2 Institute for Advanced Simulation (IAS-6), Research Center Juelich, 52425 Jülich, Germany 3Department of Computer Science, Aarhus University, 8200 Aarhus N, Denmark 4JARA Institute Brain Structure Function Relationship, Research Center Juelich, 52425 Jülich, Germany 5 Theoretical Systems Neurobiology, RWTH Aachen Univ., Aachen, Germany
*Email: p.broehl@fz-juelich.de Introduction
A typical feature of neurons in the motor cortex of mammalian brains is that they are tuned to a particular direction of movements, i.e. they exhibit most spikes when a body part is moved in a particular direction, called preferred direction (PD). It has been reported that the distribution of preferred directions among motor cortex neurons depends on the constraints in the movements: when the arm may move freely in 3D, it is uniform [1,2], but when it is constrained to a 2D movement, it is bimodal [3,4]. In this work, we aim at revealing the neuronal mechanism underlying the emergence of a bimodal PD distribution, by studying an artificial network of spiking neurons trained to control a biomechanically realistic arm model. Methods
Our model is implemented in Tensorflow [5] and consists of 300 recurrent leaky integrate-and-fire neurons with 6 linear readout neurons that control the 6 muscles in a biomechanical arm model [4]. We train it to output muscle activation signals to perform a 2D reaching task. We study its output space by applying a Principal Component Analysis on the outputs and relate the directions in this space to the directions in the joint angle arm acceleration space via Canonical Correlation Analysis. We also study the effect of each recurrent neuron on the output dynamics by interpreting its outgoing connection weights as a direction in the space of the recurrent dynamics and projecting this direction onto the output space via the readout weights. Results
The model neurons show directional tuning with bimodally distributed PDs. The output dynamics of the model are well captured by the first two principal components (PCs). The first PC aligns to the two opposite directions in the joint angle acceleration space which agree with the hand movement directions corresponding to the peaks of the bimodal PD distribution. The effects of neurons on the output dynamics concentrate around these directions. Connections between neurons with similar output effects tend to be strongly excitatory. Taken together, the core architecture of the recurrent network is characterized by two clusters of neurons with strong excitatory connections in each cluster. Connections between the clusters are mostly inhibitory. Discussion
The analysis shows that two mutually inhibiting clusters of excitatory connections underlie the control of the biomechanical arm model in a 2D reaching task by a recurrent network of spiking neurons. Since each of the two clusters is composed of neurons with similar output effect directions, which we have shown to be related to hand movement directions, the existence of the two clusters naturally explains the bimodality of the hand movement PD directions. This leads to the question whether similar structures are employed in the mammalian brain to control movements, which would be subject to future research.
Acknowledgements
This work was partially performed as part of the Helmholtz School for Data Science in Life, Earth and Energy (HDS-LEE) and received funding from the Helmholtz Association of German Research Centres. This research was partially funded by the NRW-network 'iBehave', grant number: NW21-049.
