Departmental Colloquium
- Title
- Cusp Catastrophes and Neural Computing
- Guest Speaker
- Prof. Andrew Sornborger
- Guest Affiliation
- Department of Mathematics, UC Davis
- Host
- Prof. Uwe Happek
- When
- Thursday, April 27, 2017 3:30 pm - 4:30 pm
- Location
- Physics Auditorium (202)
- Details
- The literature on information processing in neural systems is dominated by a focus on learning. Recent results have shown amazing computational abilities, in many cases with superhuman performance. Computers have been able to learn to play Atari games and Go better than we can. These successes have been based on so-called deep networks, caricatures of real, spiking neuronal networks. The work that I will present will focus on a different, but important topic in neural information processing: how the brain dynamically routes information from process to process. This ability may be seen in our ability to rapidly switch between different lines of thought. Line attractors in neural networks have been suggested to be the basis of many brain functions involving the propagation of spiking rate amplitudes, such as working memory, oculomotor control, head direction, locomotion, and sensory processing. I will discuss how, by incorporating pulse gating into feedforward neuronal networks, the propagation of information may be controlled and dynamically routed. I will show that spiking rate amplitude transmission in pulse-gated networks is associated with the existence of a cusp catastrophe, and that the slow (ghost) dynamics near the fold of the cusp underlies the robustness of an approximate line attractor that enables faithful amplitude propagation. Finally, I will demonstrate how pulse gating in combination with Hebbian learning can be used to construct predictive neural circuits and how such circuits generate oscillations with a characteristic frequency spectrum.