1
00:00:00.069 --> 00:00:7.290
Okay, so this is the Low/High Index of Pupillary Activity.
I'll try to do this in three minutes.

2
00:00:8.370 --> 00:00:14.640
It's work with myself Krzysztof Krejtz, Nina Gehrer in Germany,
Tanya Bafna and Per Baekgaard in Denmark.

3
00:00:15.960 --> 00:00:17.370
We're supported by the NSF in the US, Bevica Fonden in Denmark,

4
00:00:18.450 --> 00:00:24.390
and Polskie Ministerstwo Nauki i Szkolnictwa Wyższego in Poland.

5
00:00:25.500 --> 00:00:37.740
So in essence what this is about is the low frequency/high frequency ratio
of the pupil diameter oscillation to estimate cognitive load.

6
00:00:38.910 --> 00:00:45.180
And this is in part based on Peysakhovich's result, where he did it a Toulouse
n-back task and could differentiate between task difficulty.

7
00:00:46.710 --> 00:00:49.050
And also, in part it's based on our 2018 CHI paper on the IPA,

8
00:00:50.250 --> 00:00:56.730
the Index of Pupillary Activity where we use wavelets,

9
00:00:57.780 --> 00:01:03.870
the wavelet coefficient here gives us the detail signal and
the scaling function gives us a multi resolution analysis of the signal.

10
00:01:05.250 --> 00:01:7.770
Basically just a simple convolution using

11
00:01:09.030 --> 00:01:11.340
wavelet filters or averaging filters.

12
00:01:12.660 --> 00:01:20.490
The Haar filter, for example, in Table 4 on the right or Daubechies-4;
we ended up using symlet-16.

13
00:01:21.330 --> 00:01:28.350
The other thing that we do is find edges in the signal and that is done via
modulus maxima detection.

14
00:01:28.350 --> 00:01:34.380
So the value in the center has to be greater than or equal to both neighbors
and has to be strictly greater than one of them.

15
00:01:35.820 --> 00:01:41.400
The code for the IPA was available in the CHI paper of 2018 and we also made
it available for 2020.

16
00:01:42.630 --> 00:01:53.850
Essentially what we do is we just now take inside the wavelet transform,
we take the low to high frequency ratio. And then we compute that aspect of it

17
00:01:54.780 --> 00:01:58.780
to give us the L-H-I-P-A or LHIPA signal.
And so here we've got three experiments that we did.

18
00:01:58.780 --> 00:02:03.780
One that we replicated, or actually re-analyzed again the data from 2018,

19
00:02:04.560 --> 00:02:15.060
which was a simple fixed gaze experiment following Siegenthaler et al.

20
00:02:16.020 --> 00:02:22.230
And so those results for the 2018 paper showing that it was difficult
to count backwards by 17 or easy to count forward by two,

21
00:02:22.230 --> 00:02:28.230
and LHIPA on the right shows us good differentiation between difficult task
and the control easy, just like the IPA.

22
00:02:29.760 --> 00:02:38.520
Second experiment was no longer fixed-gaze, we followed the Tuebingen group's
design for an n-back task and we use the 1-back and 2-back

23
00:02:39.330 --> 00:02:46.160
with a limited vocabulary and this is the setup for the experiment.

24
00:02:46.160 --> 00:02:53.160
Results also showed that LHIPA can distinguish between the baseline and the
n-back tasks; the IPA could not however.

25
00:02:54.810 --> 00:03:00.960
Finally, our third experiment is an eye-typing task also easy and difficult
based on the LIX readability score.

26
00:03:02.190 --> 00:03:07.320
Here's the setup where somebody's typing a phrase, they had to remember
previously.

27
00:03:08.520 --> 00:03:14.300
And again, LHIPA shows good results differentiating between easy and difficult
tasks, but the IPA does not.

28
00:03:14.520 --> 00:03:21.300
One thing to remember is that LHIPA is inversely proportional to task
difficulty and that, in essence, is it.

29
00:03:22.350 --> 00:03:28.000
So essentially the LHIPA looks like a pretty robust way to measure cognitive
load using pupil diameter oscillation,

30
00:03:28.000 --> 00:03:34.000
whereas the 2018 IPA seems to have faltered when gaze started moving instead
of being a fixed-gaze task.

31
00:03:34.000 --> 00:03:39.630
I welcome any questions, if there are any. Thank you very much.