Project description
Motor cortex activity specifies the kinematic properties of voluntary arm movement. A decade of work has supported the “dynamical systems” hypothesis that these properties, such as movement direction, are encoded by the trajectories of population activity. But unknown is how motor cortex encodes the speed of arm movement.
Evidence exists for two competing hypotheses. In the axis hypothesis there is a dedicated speed axis: changes in the trajectory of population activity along that axis change the speed of movement. We recently showed strong evidence for what we call the traversal hypothesis: that the speed of movement is controlled by how quickly the trajectory of activity unfolds. Which hypothesis is correct is fundamental to our understanding of how mammalian cortex controls movement, and has implications for the design of brain-machine interfaces to restore movement. But these hypotheses have not been evaluated against each other nor derived predictions for the circuits that support them. In this project we will pit the two hypotheses against each other.
We will test explanations for why evidence of both hypotheses exists. To do so, we will train recurrent neural network (RNN) models to have dynamics that implement each of the coding hypotheses. With these we can test whether one coding scheme arises as a consequence of the other; and we can test whether one coding scheme arises as a consequence rather than cause of behaviour, for example reflecting posture rather than movement. Predictions of both analyses will be checked in cortical activity data from non-human primates making variable arm movements.
We will also derive the predictions that each hypothesis makes about the cortical circuit implementing it. To do so, we will train RNNs to perform a range of arm reaching tasks with varying movement speeds. Examining the resultant circuits and the dynamical systems they implement will make predictions for the properties of the motor cortex circuit and its response to perturbations that would further distinguish the two hypotheses experimentally. To achieve this we will deploy modern tools from applied mathematics to open the black box of the trained RNNs.
We will use and develop https://link.springer.com/article/10.1007/s12559-019-09634-2 recent techniques for extracting network attractors directly from the trajectory of a neural network while solving tasks. In this way we will translate the RNN realisation of coding and computation into an alternative phase space picture where transient dynamical regimes and connections between states (such as heteroclinic cycles) come to the fore. Moreover, understanding how computations are implemented in this language means that realisations in other physical networks, and notably human motor cortex, can be considered. In this way the project will develop an iterative strategy, cycling through neuroscience-AI-mathematics, for the advancement of understanding how motor cortex controls arm movement.
This is a joint project between the Schools of Psychology and Mathematical Sciences.