Title: Single-Trail Analysis of EEG during Rapid Visual Discrimination: Enabling Cortically Coupled Computer Vision
1Single-Trail Analysis of EEG during Rapid Visual
Discrimination Enabling Cortically Coupled
Computer Vision
- Authors
- Paul Sajda, Adam D. Gerson, Marios G.
Philiastides - Dept. of Biomedical Engineering, Columbia
University, USA - Lucas Parra
- Dept. of Biomedical Engineering
- City College of New Work
- USA
2Outlines
- Introduction
- Linear Methods for Single-Trail Analysis
- EEG Correlates of Perceptual Decision Making
- Identifying Cortical Process leading to Response
Time Variability - EEG-Based Image Triage
- Conclusion
3Introduction
- Linear discrimination of multichannel
electroencephalography (EEG) has used for
single-trail detection of neural signatures of
visual recognition events. - Demonstrate relating neural variability to
response variability - Studies for response accuracy
- Response latency during visual target detection
- Example Running in the park with your head
phones on, listening to your favorite tune,
concentrating on your stride, you look up see a
face that you immediately recognize as a high
school friend.
4Introduction (Con..)
- She is wearing a hat, glasses has aged fifteen
years since you last saw her. - You she are running in opposite directions so
you only see her for a fleeting moment, yet you
are sure it was her. - Your visual system has just effortlessly
accomplished a feat that has thus far baffled
best computer vision systems. - Such ability for rapid processing of visual
information is even more impressive in light of
the fact that neurons are relatively slow
processing elements to digital computers, where
individual transistors can switch a million times
faster than a neuron can spike.
5Introduction (Con..)
- Noninvasive neuroimaging has provided a means to
peer into brain during rapid visual object
recognition. - Analysis of trial-averaged event-related
potentials in EEG has enabled us to assess speed
of visual recognition discrimination in terms
of timing of underlying neural process. - Recent work has used single-trial analysis of EEG
to characterize neural activity directly
correlated with behavioral variability during
tasks involving rapid visual discrimination. - Components extracted from EEG can capture neural
correlates of visual recognition
decision-making processes on a trail-by-trail
basis.
6Introduction (con..)
- Noninvasive BCI paradigms
- Having a subject consciously modulate brain
rhythms - Having a subject consciously generate a motor
plan/visual imagery - Directly modulating subjects cortical activity by
stimulus frequency (SSVEP) - Exploiting specific ERPs (novelty/oddball P300)
- Focus Single trail detection of ERPs
relationship to visual discrimination/recognition - Cortically coupled computer vision visual
processor performs perception recognition - EEG interface detects result (decision).
7Linear Methods for Single-Trail Analysis LDA
- Conventional paradigm of evoked response
considers neuronal activity following
presentation of a stimulus. - Analyzing EEG activity of multiple electrodes
following presentation of an image. - Aim identify one type of event, visual target
recognition, Differentiate from other - Binary classification based on temporal spatial
profile of potential evoked following stimulus
presentation.
8Linear Methods for Single-Trail Analysis (Con..)
- EEG activity following each stimulus is recorded
as DxT values. - D of channels
- T of samples
- Record 1000 Hz, 64 channels
- ½ sec time window one word acquire 32000
samples - Train classifier Larger feature vector than N
100 trials - Low SNR
- Brute-force classification fail (32000-dim
feature vector)
9Linear Methods for Single-Trail Analysis (Con..)
- Exploit prior info
- S1 Reduce trial-to-trail variability by
filtering (60 Hz ) slow drifts (lt0.5 Hz), no
info carry - S2 Reduce dimensionality by grouping signal into
L sample blocks does not change within time
window. - S3 Increase of trails by L redundant
samples/window - L50 signal of interest at 10 Hzgtgt signal noise
- Transform original data for each trail with TD
dimensions into L trails TD/L - L50, N100, training exam5000
- train classifier with 640-dimensional feature
vector at 10 Hz
10Linear Methods for Single-Trail Analysis LDA
(Con..)
- linear classifier TD/L dimension, good results
- Single training window LT, D-dimensional
feature vector - Linear Classification
- Feature vector x is projected onto an orientation
defined by vector w - Projection, y wTx, optimally differentiates
betn two classes.
11(No Transcript)
12Linear Methods for Single-Trail Analysis (Con..)
- Adv
- linear voltage combination-immediate
interpretation (current) - Coefficients that coupled current with observed
voltages given for linear model by altxTygt/lty2gt - Angular bracket indicate average over trials
samples. - Coefficients a describe coupling ( correlation)
of discrimination component y with sensor
activity x. - a x D-dim vector(row/column)
- Strong coupling low attenuation visualized
intensity map-sensor projections
- Easy to implement
- Fast
- Permit real-time operation
- Disadv
- does not capture synchronized activity gt10 Hz
- Neither does it capture activity that is not a
fixed distance in time from stimulus - Only phase-locked activity detection
13EEG Correlates of Perceptual Decision Making
- Single-trail LDA to identify cortical correlates
of decision-making during rapid discrimination of
images. - Experiment Psychophysical performance is
measured for several subjects during RSVP task - A series of target (faces) and non-target (cars)
trials presented in rapid succession (Fig. 25.2a) - Simultaneously recording neuronal activity from
64-channel EEG electrode array. - Stimulus evidence varied with phase coherence
(Fig. 25.2b)
14EEG Correlates of Perceptual Decision Making
(Con..)
- Within trail block face car images over a
range of phase coherence presented random order. - 12 face 12 car, grayscale images (512x512, 8
bits/pixel) - Equal of front side views
- Equal spatial frequency, luminance contrast
- Subjects are required to discriminate
- type of image (face/car)
- report decision by pressing a button.
15EEG Correlates of Perceptual Decision Making
(Con..)
- EEG data acquired in Electrosatically shielded
room - Used Sensorium EPA-6 Elecrophysiological Amp.
from 60 Ag/Agcl scalp electrodes, 3 periocular
electrodes placed below left eye, left right
outer canthi - Sample rate 1000Hz, 0.01-300 Hz passband, 12
dB/octave HPF 8th order elliptic LPF. - Software-based 0.5 Hz HPF (remove dc drifts)
- 60/120 Hz notch filter (minimize line noise)
- Record EOG signals, remove motion blink
- Motor response stimulus events recorded on
separate channels
16EEG Correlates of Perceptual Decision Making
(Con..)
- Identify EEG components that maximally
discriminate betn 2 conditions - Identify 2 time window for discrimination
- Identify early (?170 ms following stimulus)
late component (gt300 ms) - LD trained by integrating data across both time
windows (2D-feature vector) - Performance improved (higher Az)
- Fig. 25.3 comparison of psychometric
neurometric functions for 1 sub in dataset. - All subjects a single function can fit behavioral
neuronal datasets 2 sep functions.
17EEG Correlates of Perceptual Decision Making
(Con..)
- Data analyzed both stimulus response-locked
conditions - Both face selective components appear to be more
correlated with onset of visual stimulation than
response for one subject (Fig. 25.4) - Discriminate activity significant down to 30
phase coherence (early late component) - Randomize Az significant level of plt0.01.
- Results neural correlates of perceptual decision
making identified using high-spatial density EEG
corresponding component activities
18EEG Correlates of Perceptual Decision Making
(Con..)
19Identifying Cortical Process leading to Response
Time Variability
- Significant variability in response time observed
across trials in many visual discrimination
recognition. - Factors difficulty in discriminating an object
on any given trial - Trial-by-trial variability of subjects
engagement - Intrinsic variability of neural processing
- Identifying neural activity correlated with
response time variability may shed light on the
underlying cortical networks responsible - for perceptual decision-making processes
processing latencies that these networks may
introduce for a given task
20Identifying Cortical Process leading to Response
Time Variability (Con..)
- Visual target detection use RSVP single trial
spatial integration of high-density EEG - to identify time course cortical origins
leading to response time variability. - RSVP paradigm (Fig 25.5)
- Varied scale, pose position of target objects
(people) requires subjects to recognize rather
than low-level features - Continuous sequence of natural scenes
- 4 blocks (50 sequences) rest period no more than
5 mins.
21Identifying Cortical Process leading to Response
Time Variability (Con..)
- Each sequences 50 images, 50 chance containing
one target image with one or more people in a
natural scene. - Target appear middle 30 images/50 sequences
- Detractor image remaining natural scene without
a person - Each image 100 ms
- Fixation cross 2s betn sequences.
- Instructed to press left button of a 3-button
mouse with right index finger while fixation
cross present, release as soon as recognize
target image. - Used LD to determine spatial weighting
coefficients that optimally discriminate betn
EEG resulting from different RSVP task conditions
(target vs. distractor) over specific temporal
window betn stimulus response.
22Identifying Cortical Process leading to Response
Time Variability (Con..)
- Integration across sensors enhances signal
quality without loss of temporal precision common
to trail averaging in ERP. - Resulting discrimination components describe
activity specific to target recognition
subsequent response for individual trails. - Intertrial variability estimated by extracting
feature from discrimination components. - Peaks of spatially integrated discriminating
components found by fitting a parametric function
to extracted component y(t)
23Identifying Cortical Process leading to Response
Time Variability (Con..)
- Gaussion Profile parameterized by its height ?,
width?, delay ? and baseline offset ? - Response-locking of discriminating components
determined by computing linear regression
coefficients that predict latency of component
activity, measured by ? from response time given
by r. - ?j ?rj b ?j-peak latencies, rj-response
time for j-th trial, b-offset term for
regression (Fig. 25.6) - Proportionality factor from response time peak
latency regression (?) degree of
response-locking () for each com. - Quantify extent to which component is correlated
with response across trails.
24- 0 pure stimulus
- 100 pure response lock
- ?100 slow responses late activity
- fast responses
- early activity
- ?0 timing activity does not change with
response time therefore stimulus locked.
25Identifying Cortical Process leading to Response
Time Variability (Con..)
- Group results discriminating component activity
across 9 participants (Fig. 25.7) - Scalp projections normalized prior averaging
- A shift of activity from frontal to parietal
regions over the course of 200 ms. - Scaled response times and component peak time are
concatenated across subjects - These registered group response times then
projected onto scaled component peak times to
estimate degree of response-locking across
subjects. - Group response lock increased from 28 (200ms) to
78 (50 ms) after response.
26 27Identifying Cortical Process leading to Response
Time Variability (Con..)
- Significant processing delay introduced by early
processing stages - Within 200 ms prior to response (250 ms following
stimulus), activity is already, on average, betn
25 35 response-locked - It is possible that some of this early
response-locking may be due to early visual
processes (0-250 ms post stimulus - For 9 subjects correlation Analysis reveal that
Discrimination Component activity progressively
becomes more response-locked with subsequent
Processing stages
28EEG-Based Image Triage
- EEG system capable of using neural signatures
detected during RSVP to triage sequences of
images, reordering them so target images are
placed near the beginning of sequence. - Called Cortically coupled computer vision
- Robust recognition capabilities of human visual
system (invariance to pose, lighting, scale) - use a noninvasive cortical interface (EEG) to
intercept signatures of recognition events- - visual processor performs perception
recognition EEG interface detects the result
(decision) of processing
29Image Triage (Con..)
- Experimentation Series of self-paced feedback
slides were presented indicating position of
target images within sequence before after EEG
triage. - Participants completed 2 blocks (50 sequences)
with brief rest period lasting no more than 5
minutes betn blocks - During 2nd block participants were instructed to
quickly press left button of 3-button mouse with
right index finger as soon as they recognized
target images. - Button press twice if one target immediately
followed other. - Not respond with a button press during 1st block.
- Classify EEG online used Fisher linear
discriminator - to estimate a spatial weighting vector that
maximally discriminate between sensor array
signals evoked by target non-target images.
30Image Triage (Con..)
- 5000 images presented to subject in sequences of
100 images (with without motor response) - Training 2500 images (50 targets, 2450
non-target) - Training window 400-500 ms following stimulus
onset was used to extract training data - Weights updated adaptively with each trail during
training - Classification threshold adjusted to give optimum
performance for observed prevalence (Class-prior) - These weights threshold were fixed at end of
training period applied to subsequent testing
dataset (images 2501-5000) - Offline triage after experiment multiple
classifier with different training window onsets
were used.
31Image Triage (Con..)
- Training window 0900 ms in steps of 50 ms
- Duration 50 ms
- After trained, optimal weighting outputs found
using logistic regression to discriminate target
non-target images. - EEG data evoked by 2500 images to train find
inter-classifier weights. - applied to testing dataset evoked by 2nd set 2500
images (25015000) - Following experiments all image sequences were
concatenate to create training testing
sequences that each contain 2500 images (target
50, non-targets 2450). - Image sequences are restored according to output
of classifier with multiple training windows for
EEG evoked by every image.
32Image Triage (Con..)
- Comparison Sequence triaged based on button
response. - Image restored
- RT-onset of button response (occurs within1 sec)
- P(target/RT)0, no response
- Priors P(target) 0.02
- P(non-target)0.98
- P(RT/target) Gaussian distributions with mean
variance determined from response times from
training sequences - If target appeared 1st in sequence 2 button
response occurred within 1 sec-assigned to target
images - 2nd response to 2nd target image
33Image Triage (Con..)
- Testing sequences if 2 or more within 1 sec,
response with greatest p(target/RT) assigned to
image. - P(RT/non-target) is a mixture of 13 Gaussians
(same variance as p(RT/target)) - Means shifting mean from p(RT/target) 600 ms in
past to 700 ms in future, increments of 100 ms,
exclude actual mean of p(RT/target) - Triage results (subject 2) Fig. 25.8
- of targets as a function of of distracter
images both before after triage based button
press EEG. - area under curve generated by plotting fraction
of targets as a function of fraction of
distracter images presented is used to quantify
triage performance.
34(No Transcript)
35Image Triage (Con..)
- Area is 0.50 for all unsorted image sequences
since target images are randomly distributed
throughout sequences. - Ideal 1.00
- No significant diff in performance (button-based
EEG-based) triage (0.93?0.06, 0.92 ? 0.03,
p0.69, N5) - EEG (motor no-motor)(0.92?0.03, 0.91 ? 0.02,
p0.81, N5) - position of target images (black squares)
non-target images (white squares) in images
sequences (Raster 25.8b-f) - Performance (5 subjects) Table 25.1
36Image Triage (Con..)
- Performance (5 subjects) Table 25.1
- High-level performance
- Button-based triage performance begins to fail,
when subjects do not consistently respond to
target stimuli (response timegt1 sec), subject 2
only 74 - EEG (motor button) increase triage performance
37Conclusion
- Invasive noninvasive EEG recording obtained
during RSVP of natural image stimuli have shed on
speed, variability spatiotemporal dynamics of
visual processing. - Future issues learning/priming/habituation,
effect of subject expertise, image type
category - Application level intercepting neural correlates
of visual discrimination recognition events
that effectively bypass slow noisy motor
response loop.
38Thanks