Jonathan Timcheck
Theory: Spiking Networks

Personal Background

  • BS, Engineering Physics, Ohio State University, 2015
  • MASt, Applied Mathematics, University of Cambridge, 2016
  • PhD, Physics, Stanford University, co-advised by Surya Ganguli and Kwabena Boahen (in progress)
Jonathan Timcheck

Research Goals

The brain is a magnificent computational device — we can dance, read, and learn — but how does the brain achieve these computational feats? Intriguingly, the brainís architecture is vastly different than that of todayís digital computers. The brain computes in a distributed, massively parallel fashion with a noisy mess of recurrently connected neurons, whereas digital computers execute instructions step-by-step with a highly organized arrangement of pristine logic gates. This stark contrast suggests that we have a lot to learn from the brainís style of computation if we are to design computers on par with the brain; I seek to understand the brainís computational strategies.

Project Status

At face value, one might expect the neuronal noise in the brain to hinder computation, but counterintuitively, it turns out that it may in fact benefit computation. I study this idea in spiking neural networks synthesized using the predictive coding framework (Deneve et al.). In this framework, neurons that code for a similar quantity instantaneously communicate with each other so that only one neuron spikes to represent the quantity; this results in an efficient code. However, realistically these communications cannot be instantaneous due to biological spike propagation delays, and so extra neurons may spike before they receive the othersí communication; this destroys the codeís efficiency. By introducing noise, one can diversify the neuronsí anticipated spike times, thus allowing enough time for communication and preventing the extraneous spikes. On the other hand, introducing too much noise destroys the fidelity of the representation. I work to quantify this fascinating tradeoff and determine the non-zero noise level that yields optimal performance.