Information transfer rate · Lobeworks/17

The information transfer rate is the number of bits a brain-computer interface delivers per selection or per minute, computed from how many options the user chooses among, how often the choice is right and how long each choice takes, and it is the figure used to compare interfaces that work in different ways: a speller

Information transfer rate. The information transfer rate is the number of bits a brain-computer interface delivers per selection or per minute, computed from how many options the user chooses among, how often the choice is right and how long each choice takes, and it is the figure used to compare interfaces that work in different ways: a speller, a cursor, a yes-or-no switch.

The usual formula is Wolpaw's. For NNN options chosen with accuracy PPP, each selection carries

B=log⁡2N+Plog⁡2P+(1−P)log⁡21−PN−1B = \log_2 N + P \log_2 P + (1-P)\log_2 \frac{1-P}{N-1}B=log2​N+Plog2​P+(1−P)log2​N−11−P​

bits, and multiplying by selections per minute gives a rate. A yes-or-no switch right 95 times in 100 carries 0.71 bits per selection, and one right 99 times in 100 carries 0.92.

The formula assumes every option is equally likely and every error equally likely to land on any wrong option. Real interfaces break both, so the number is an idealisation, and it tracks poorly with what users actually achieve.

Accuracy hides imbalance. When almost every request deserves a yes, a system that always says yes scores high accuracy and misses every refusal, which is why uses where a refusal matters report false authorisations and missed stops on their own.

It measures the channel. Measures of utility, which count the time needed to reach a correct result with the corrections included, follow what the user gets more closely.

Questions: Why do speech prostheses report words per minute and error rates instead of bits per minute? The bit rate assumes a fixed set of equally likely options, and speech has neither: a vocabulary of a hundred thousand words, used with very unequal frequencies, and a language model that makes some sentences far more expected than others. Words per minute says how fast a person talks, the word error rate says how much of it was wrong, and the two together describe the trade a user lives with. A rate quoted without its error rate describes half of it.