Neural network properties for 3D reach planning

G. Blohm1,3, G. P. Keith2,3, J.D. Crawford2,3

1 Centre for Neuroscience Studies, Queen’s University, Kingston, Ontario, Canada
2 Centre for Vision Research, York University, Toronto, Ontario, Canada
3 Canadian Action and Perception Network (CAPnet)

Planning a reaching movement based on visual information of the hand and movement goal location requires two important steps. First, the hand and goal locations must be compared to produce a desired movement vector. Second, the initial visual coordinates of hand and target location ultimately need to be converted into motor coordinates appropriate to move the arm. It remains largely unknown how the 3D aspects of this could be implemented in the brain and what neural properties to expect in areas involved in these computations. Here, we hypothesize that both computations are performed in parallel, making use of distributed computing principles in neural networks.
To analyze the emerging neural properties that one might expect to find in areas involved in these computations, we trained a 4-layer feed-forward artificial neural network to plan a 3D reach vector from visual inputs of hand and reach goal. We also included 3D extraretinal eye and head orientations to specify body geometry. This was crucial to dissociate translational and rotational geometry, without which apparent reference frames of the network’s units cannot be determined.
Receptive field gradient analysis showed that the computation of the motor vector mainly emerged at the level of the 3rd layer in the network (presumably analogous to pre-motor cortex), while the 2nd layer (hypothetically equivalent to posterior parietal cortex) mostly encoded hand and reach goal locations independently. Comparing receptive fields, motor fields and simulated micro-stimulation, we found that individual units within and across layers can display different reference frames across all three techniques, even within a single unit. We interpret this finding as evidence for partial reference frame transformations at the singe-unit level. The extensive presence of gain-like modulations of unit activity with changes in eye and head orientation further suggests that gain mechanisms could be used to combine these partial transformation modules in a purposeful fashion to produce the required overall transformation at the population level.
Through comparison of our analysis with data in the literature, we suggest rationales for previously contradictory or unexplained findings, e.g. why there should be gaze-centered receptive fields in pre-motor cortex. In addition, we make predictions as to what one should find when using different electrophysiological techniques to probe neural properties in a brain area involved in 3D reach planning. We believe that this new theoretical framework could lead to a better understanding of the neuronal mechanisms involved in complex sensory-motor mechanisms beyond reach planning.

Supported by: Marie Curie (EU), CIHR (Canada)