3 ms·
One thing that concerns me a bit is stacking machine learning onto this. Imagine a 100% paralyzed person in the hospital. We attach some probes to their head an
by CobrastanJorji 4y ago
One thing that concerns me a bit is stacking machine learning onto this. Imagine a 100% paralyzed person in the hospital. We attach some probes to their head and get a very noisy signal. What comes back is a very noisy signal, but we've got a machine learning model that takes in a huge amount of context about interacting with computers and text and such and uses it as part of how it cleans up the signal, just like how phones today use machine learning to de-noise camera input.
Then let's say that our paralyzed person says something very impactful like "please turn off the life support" or "I will all of my money to my kids." Can you imagine the lawsuit over to what degree the machine learning was driving the boat?
- ggm 4y agoRequire them to spell it out letter by letter. Which btw is almost certainly how current Brain reading works, you think hard to stop a pointer walking a screen.
- rhacker 4y agoOr just ask a lot of questions which prove mindfulness.
- mouse_ 4y agoHow can you design questions that can't be answered by an LLM?
- Apocryphon 4y agoAsk them something specific that only they would know. Classic anti-cold reading technique.
- ggm 4y agoDumas has fun with this in "the Count of Monte Cristo"
- KRAKRISMOTT 4y agoA lot of them are viterbi estimators, you lose predictive power when you do character level only.
- jrumbut 4y agoMost systems I have seen allow for you to delete characters. This was close to 20 years ago now though so perhaps it has gone through a worse is better phase. Now, the issues you state are very real, terrifying, and depressing. Still anything that allows the person to communicate directly, rather than through an intermediary holding an eye board, will be a major step toward greater independence and the ability to hold those who abuse the disabled accountable.
- a-dub 4y agoin research, typically the behavioral tasks are randomized to avoid this sort of thing. for practical purposes and legal decision-making, i have always envisioned a protocol where the patient is required to repeat back a random string in order to authenticate the system's output. that is, the patient is shown or played a recording of the authentication sequence (by a separate device) and then they must express it through the bci before any device output from that session can be used as communication of intent for legal purposes.