Text

theses

ETD doctoral

We present a data-driven computational approach for studying neural systems. In this approach one starts with experimental stimuli (inputs) and measured neuronal responses (outputs). The relationship between the inputs and outputs is modeled with an artificial recurrent neural network (ARNN). A detailed investigation of the network weights and response properties of the connected elements, together with simulated experiments performed on the ARNN leads to significant new insights and new hypotheses about the underlying neural mechanisms. We first applied this approach to motion responses of neurons in the macaque middle temporal area (MT). This provided the novel insight that recurrent networks dynamics can explain complex motion tuned response dynamics found in MT neurons, without the need for feedforward temporal delay lines. In our second study we used this approach to model the early visual form processing pathway of the macaque brain. Neurons in the secondary visual cortex (V2) were stimulated with textured stimuli designed to probe the visual systems for complex visual shapes. The approach led to the novel hypothesis that selectivity for complex form depends on selectivity for motion. For the third study we extended the approach by taking advantage of chronically implanted microelectrode arrays (FMA) in primary visual cortex (V1) of the awake behaving macaque. With the FMA we collected V1 responses on day one, fitted an ARNN, explored the detailed properties of the ARNN the following days, and tested model predictions with a V1 validation experiment within the same week. We found that V1 selectivity for form is much more complex than commonly thought and includes spatiotemporal interactions between multiple hotspots in the receptive field. With this approach we found complex V1 tuning properties that are currently thought to primarily arise higher up in the visual processing stream. We conclude that ARNNs can offer a useful tool set for systems neuroscience; the powerful computational approach, together with carefully designed experiments, provides novel hypotheses and insights into the complexity of neural function.

ETD_7457

doi:10.7282/T32809ZR

Ph.D.

Includes bibliographical references

by Jeroen Joukes

The author owns the copyright to this work.