Great paper: a brain-computer interface lets a patient "type" at ~90 char/min (avg smartphone typing: ~110c/m) by imagining themselves handwriting biorxiv.org/content/10.110 ht
Even better—they used a ~simple ML arch and very little training data! Low-hanging fruit?
Conversation
Micro-electrode arrays are invasive so not sure if it's low enough for low hanging fruit :)
Still very cool, I like how the T-SNE viz shows how similar the signals are per letter!!
1
10
Indeed that doesn't look low enough for me to try ;)
2
Facebook announced a similar goal of ~100 wpm via BCI a few years back
Yeah, all invasive brain imaging has a huge barrier to overcome. A bigger problem in that area would be how to overcome rejection by the immune system. Compared to it, signal analysis and machine learning is easy.
4
I suspect they tried with subjects imagining saying the sentences, so it appears imagining writing delivers better results than speaking? Interesting if true.
1
1
“Speaking” or imagined thought is far more distributed - inner speech decoding has had interesting attention recently with ECoG (see knightlab.berkeley.edu/statics/public) and (changlab.ucsf.edu/publications/s)
1
4
super-fascinating tho in my rapid skimming there's no cognitive level theory. One of my interests is using sleep onset-like information processing, including motor imagery, to induce sleep (serial diverse imagining, cognitive shuffling)
2
in that case the output of motor imagery interacts (somehow) with lower levels of the SO control system
It's going to be very interesting to watch demand once it outperforms any keyboard user.