Adaptive event-driven simulation strategies for accurate and high performance retinal simulation

The biological retina is a complex system with intricate neural processing, dense connectivity and massive number of neurons. Unlike most neurons, most retinal neurons respond with graded potentials that will increase the retinal system neural-state simulation update rate. The retinal neurons connect densely to each other using various connectivity patterns including feedforward, feedback and laterally leading to an even higher neural-state update rate. Considering also the massive number of neurons, modeling the retinal system is very challenging.;The lack of medical cures for common retinal disorders has motivated many research groups to study the retina and to attempt to model this complicated system. Some of the retinal projects have created retinal prosthesis devices fabricated on chips, while other groups have created retinal simulators. Retinal simulators are used as powerful learning tools to understand the retinal neurons and their role in vision. For retinal prosthesis research groups, retinal simulators are important tools to determine a suitable quantity of neurons in the prosthetic device design (e.g. number of photoreceptors to allow face recognition).;We describe an efficient event-driven retinal simulator that speeds up the simulation by detecting periods of "inconsequential activity'' in the neural responses, and carries out approximations or eliminations of simulation events during these periods. We implemented several simulation acceleration algorithms including a lookahead event-driven algorithm that carries out the approximations and self-regulating and event-merging event-driven algorithms that eliminate events.;We conducted various simulation experiments using a variety of inputs and showed the acceleration and error trade-offs for each input. The performance of each algorithm has been computed relative to the baseline event-driven simulation that does not include any acceleration algorithms.;The lookahead event-driven algorithm carries out a piecewise linear approximation to approximate the future response of the node being simulated given its last two responses. In the lookahead event-driven algorithm we propose, events that are scheduled at a node ni with a delivery time less than or equal to tcurrent+TB i can be delivered to ni at t current, such that TBi is n i time bucket. The lookahead carries out an approximation of the time bucket events using the last two computed responses. Reducing computations will increase the simulation speed, but will possibly reduce accuracy. The time bucket value is determined by the modeler.;Event merging and self-regulating aim to reduce the number of events in the system to increase the simulation speed while maintaining accurate results. Self-regulating operates to determine if the simulation can drop a low-value event without processing it. A low-value event is an event that does not influence the destination node's response or the influence is weak. The event-merging acceleration algorithm attempts to create more-complex meta events, each of which entails a set of events that occurs within a time bucket. The algorithm scans the event queue at certain times and attempts to create one event object (meta event) from several events that are scheduled to be delivered to the same destination with delivery times that fall in the same time bucket.;The lookahead algorithm performed well for experiments with linearly changing responses. The self-regulating algorithm performed well for slowly changing responses since it reduces the mean number of reported events from the slowly changing nodes. And, the event-merging acceleration algorithm performed well for nodes receiving many events within short periods of time since this increases the probability of merging a larger number of events.;Our main contributions are the following: First, we describe a comprehensive outer retinal simulator framework that is flexible and biomimetic, and processes realistic input, producing tangible output. Second, the simulator employs novel event-driven simulation acceleration strategies that are capable of increasing the simulation speed at much higher rate than the error increases. These strategies would be helpful on any continuous system simulation and are not unique to retinal or biological simulations.