Paper Reading #22

Comments:
Comment #1
Comment #2
Reference:
Title: A POMDP approach to P300-based brain-computer interfaces
Author: Jaeyoung Park, Kee-Eung Kim, Sungho Jo
Venue: IUI 2009/2010
Summary:
Brain Computer Interfaces (BCI) provide a channel of communication for "conveying messages and commands from the brain to external sources." This is significant because by studying BCIs we can improve communication between the brain and the computer. Perhaps the most popular way of measuring BCIs is by the use of EEGs. EEGs are popular because they are non-invasive, cheap, and easy to use. Also, it has a very reliable signal feature called a P300, which "a positive peak in the signal amplitude at about 300ms after a stimulus is given to the user’s attention." One system that has been proposed using the P300 and EEGs is the P300 speller. The P300 speller has a 6x6 matrix of letters, and the letters randomly and systematically flash, so when the letter that the user is looking at flashes, the P300 is received approximately 300 ms later. This is the BCI system that this group used to test their algorithm.

Most BCI work has been focused on the lower-level of the interface, being feature extraction or classification methods. This paper focuses on a higher level problem: the optimal order or sequence of flashes. To do this, the group used a Partially Observable Markov Decision Processes (POMDP) algorithm. Basically, this algorithm uses several variables (8 total) and combined with the given uncertainty to model sequential decision making. To implement this, raw signals are sent to the preprocessor, which is a collection of bandpass and other filters. Then that is classified to determine if a P300 occurred. Then the system is modeled and sent to the POMDP planner.

The baseline for the experiment was BCI algorithms already in use. The group tested 2x2 and 2x3 matrices. The algorithms the group used to compare to the baseline were POMDP with select actions (PWSA) and without select actions (PWOSA). The PWSA improved the bit rates the most, by 242% to 265%. This was 135% to 151% better than the baseline algorithms already being used.
Discussion:
I found this paper mildly enjoyable. I think this is definitely a usable innovation. I admit I don't know much about BCIs, but the stats presented show that this is an obvious improvement. The thing I didn't like about the algorithm was that it can't be used on any data because it is computationally infeasible. However, this is definitely an improvement and in a useful area. The paper used a TON of mathematical notation and explanation. If you've read my blogs before, you know I don't like that in a paper like this. I think that stuff should be left for a follow up, once you've already captured the audience's attention and have them asking for more. The concept is strong, but the paper loses the reader's attention.