PMID-12052948[0] Direct Cortical Control of 3D Neuroprosthetic Devices
- actually not a bad paper... reasonable and short. they adapted the target size to maintain a 70% hit rate, and one monkey was able to floor this (reach and stay at the minimum)
- coadaptive algorithm removed noise units based on (effectively) cross-validation.
- both arms were restrained during performance & co-adaptation. Monkeys initially strained to move the cursor, but eventually relaxed.
- Changes from hand control to brain control random but apparently somewhat consistent between days.
- continually increasing performance in brain-control for both monkeys, arguably due to the presence of feedback and learning. They emphasize the difference between open-loop (Wessberg) and closed-loop control. (42 ± 5% versus 12 ± 5% of targets hit)
- still, the percentage of correct trials is low - ~50% for the 8 target 3D task.
- monkeys improved target hit rate by 7% from the first to the third block of 8 closed-loop movements each day.
- claim that they were able to record some units for up to 2 months ?? ! In their other monkey, with teflon/polymide coated stainless electrodes, the neural recordings changed nearly every day, and eventually went away.
- quote: Cell-tuning functions obtained during normal arm movements were not good predictors of intended movement once both arms were restrained. interesting.
- coadaptive algorithm:
- Raw PV yielded poor predictions.
- first, effectively z-score the firing rate of each neuron.
- junk / hash neurons were not removed.
- Two different weights per neuron per axis (hence 6 weights altogether), one if firing rate was above the mean value, another if it was below. corrected for resulting drift. Sum (neuronal firing rates * weights) controlled velocity on each of the axes. (Hence, it is not surprising that the brain-control tuning was significantly different from the hand control - the output model is vastly different).
- restarted the coadaptive algorithm every day?
- coadaptive algorithm appears to be something like stochastic gradient descent with a step-size that decreases with increasing performance.
- From her Case-western website, Dawn Taylor still seems to be on the coadaptive kick. Seems like it's bad to get stuck on one idea all your life ... though perhaps that is the best way to complete something.
- Their movies in supplementary materials look rather good, better than most of the stuff that we have done. She did not quantify SNR or correlation coefficient.
____References____
|