R.L.White*; J.T.Baker; L.H.Snyder

Neurons in the lateral intraparietal area (LIP) represent remembered spatial locations in retinotopic coordinates. During a gaze shift, the representation of a target fixed in the world (world-fixed reference frame) must be updated, while the representation of a target fixed relative to center of gaze (gaze-fixed) must remain constant. We trained a three-layered recurrent neural network to remember the location of visual targets as world-fixed or gaze-fixed based on a contextual cue and compared the network with neurophysiological data obtained from monkeys performing the same task. The neural network was able to perform reference frame dependent updating with either gaze position or gaze velocity as input. When only position information was provided, activity modulations by gaze position (gain fields) were observed in the middle (hidden) layer. When both position and velocity were provided, the network preferentially relied on velocity, and gain fields were weak or absent. Thus, a simple recurrent network architecture can compute and maintain spatial location in either of two reference frames, even without positional information or gain field mechanisms. Hidden layer nodes shared some features with single neurons in LIP. Depending on the contextual cue, all neurons and most hidden nodes appropriately compensated for or ignored a gaze shift. Furthermore, both hidden nodes and neurons showed incomplete compensation for gaze shifts on world-fixed trials, even when network output and animal behavior were accurate. Activity in LIP, like the hidden layer, may reflect a partial sensorimotor transformation that is read out by downstream structures analogous to the output layer.
Supported by: NIH; McDonnell and EJLB Foundations

Download a PDF of the poster

Citation: R.L.White, J.T.Baker, L.H.Snyder. REFERENCE FRAMES AND SPATIAL MEMORY OPERATIONS: A NEURAL NETWORK MODEL Program No. 57.17. 2002 Abstract Viewer/Itinerary Planner. Washington, DC: Society for Neuroscience, 2002. CD-ROM.