The simulation results of the neuronal firing data transmission model show that the bursting and super-bursting times present the most critical intervals for the system. In BMI applications, for example the queuing-based transmission delay must fall below 10 milliseconds. Setting a value for the transmission rate will not only depend on the limits for queuing-based transmission latency, but also on the memory resources available to save the queue on the hardware used.
The data sets recorded from the hippocampus neurons showed a synchronous multichannel bursting, and had the maximum queue depth requirements based on the simulations in MATLAB. If the maximum queue depth is 15,000 spikes and a spike waveform holds 50 samples with each sample being 2 bytes long, then the design requires a memory of approximately 1500 Kbytes just to save the queue. A memory space of 1500 Kbytes would translate to approximately 370 BRAMs (36Kbits each) on the FPGA. Following the literature recommendation, the transmission rate was set at five times the MFR. Increasing the data transmission rate would ease the memory burden on the FPGA resources, as some FPGA models do not have this amount of BRAMs. For example in the hippocampal recordings the TR was about 176Kspikes/sec. Considering 50 samples per spike and 2 bytes per sample we the TR was modeled at 17.6 Mbytes/s. Considering the data transmission options from an FPGA to a host, such as Ethernet and PCIe, there is some room for transmission rate increase. With the Ethernet and PCIe offering transmission rates in the range of at least few hundreds of Megabytes per second.
The recordings acquired from the cortical neurons demonstrated a more uniform firing rate, and a much lower average value. This fact was reflected on the results obtained for the queue depths and associated latencies.
The hippocampus is a main component of the brains of vertebrates. It is located under the cerebral cortex and belongs to the limbic system. It plays a major role in fusing the information from short-term memory to long-term memory and in spatial navigation. The subiculum, a component of the hippocampal formation, is thought to perform relaying of signals originating in the hippocampus to many other parts of the brain . In order to perform this function, it uses intrinsically bursting neurons to convert promising single stimuli into longer lasting burst patterns as a way to better focus attention on new stimuli and activate important processing circuits. The detailed explanation of the firing dynamics of neurons from different parts of the brain is beyond the scope of the dissertation, but it was appealing to search for an explanation for the reason behind the similarity between the results obtained from the hippocampal data set used in testing and the firing patterns of Intrinsically Bursting IB neurons.
Limitations of the model:
(1) The model, designed in MATLAB, assumed that the spikes within a 1msec bin are sent to the output FIFO as a block at the same instant. The data used was recorded at a sampling rate of 7.022 KHz, meaning that the bin collected the spikes occurring across a time approximately equivalent to seven sampling periods. With a high count of channels and high transmission rates the binning size of 1msec may be relatively large.
For a TR of 50 Msamples/sec, and 48 samples per spike, the queue can empty 148 spikes within a sampling period, i.e. before any new samples arrive.
(2) In real implementation, the channels are recorded using TDM, so exact synchrony will not be faced.
(3) At high channel counts, the transmission rate values increase to limit hardware memory. With the TR values approaching the range of the clock frequency on the FPGA, the bulk transmission assumption will not be accurate, and the read and write times as well as the sample by sample transmission to the FIFO have to be considered.
