The challenge isn’t just the hardware. Software in neuroscience is in an exceptionally sorry state.
As I write, I have a job running to infer the number of neurons and their spike trains from a recording done with 4096 electrodes. Sampling of voltage is done at 18kHz, and the big challenge is dealing with spikes that are observed from multiple electrodes. One approach to this is to detect spikes by a threshold and convert to a 5ms voltage vector, run PCA over these vectors, construct template waveforms for each cluster, and then merge similar clusters over neighboring electrodes.
It will take >24 hours to just extract the spike trains from a 1 hr recording, and that’s not including any interesting analysis. Fortunately I can run this on a few dozen nodes and start approaching real-time, but we are quite far off from being able to handle these data firehoses at the rate necessary for BCI.
I don't know much about BCI, but I'd imagine not all applications require spike sorting (http://www.scholarpedia.org/article/Spike_sorting for those who are unfamiliar with this), though of course if one could spike-sort in close to real time, one could do a lot more.
In any case, I absolutely agree with you -- at the algorithmic and implementation level, there are lots of nice problems in making sense of neuronal recordings.
This is a well researched and well articulated piece that largely gets it right regarding the importance of recording intracortical signals for neural interfaces. These address the significant limitations introduced by non-invasive methods like EEG.
As I write, I have a job running to infer the number of neurons and their spike trains from a recording done with 4096 electrodes. Sampling of voltage is done at 18kHz, and the big challenge is dealing with spikes that are observed from multiple electrodes. One approach to this is to detect spikes by a threshold and convert to a 5ms voltage vector, run PCA over these vectors, construct template waveforms for each cluster, and then merge similar clusters over neighboring electrodes.
It will take >24 hours to just extract the spike trains from a 1 hr recording, and that’s not including any interesting analysis. Fortunately I can run this on a few dozen nodes and start approaching real-time, but we are quite far off from being able to handle these data firehoses at the rate necessary for BCI.