Week 9 Flashcards
Describe the characteristics of the five frequency bands (or waves) of the electroencephalography (EEG) signals that are prominent in certain states of mind
1) Delta waves (1-4 Hz), associated with unconscious mind
during deep dreamless sleep
2) Theta waves (4-7 Hz): associated with
subconscious mind, e.g., during sleep/dreams
3) Alpha waves (8-13 Hz)
associated to relaxed mental state more visible over parietal &
occipital lobes, high alpha activity correlates to brain inactivation
4) Beta waves (13-30Hz)
related to an active state of mind more prominent in frontal cortex &
over other areas during intense focused mental activity
5) Gamma waves (> 30 Hz), associated
with hyper brain activity
Physiological signals can be measured using
1) Electroencephalography (EEG):
- poor spatial resolution & requires many electrodes
- provides great time resolution ⇒ to study phase changes in response to emotional stimuli
- non-invasive, fast & inexpensive
- due to wearability, price, portability & ease-of-use of
wireless EEG devices
- ⇒ diverse applications:
entertainment, e-learning, virtual worlds, e-healthcare,
instant messaging, online games, assisting therapists &
psychologists
2) Functional Magnetic Resonance Imaging (fMRI)
3) Positron Emission Tomography (PET)
EEG distance betwen adjacent electrodes?
10%/20% of total
front-back/right-left distance of skull
EEG hemisphere location identifier?
1) z (zero) - electrode placed on mid line
2) even numbers - electrode
positions on right hemisphere
3) odd numbers - electrode
positions on left hemisphere
Anatomical landmarks for correct positioning of electrodes
1) nasion - point between forehead & nose
2) inion - lowest point of skull from back of head
3) pre auricular points anterior to the ear
Two types of electrodes EEG
1) monopolar - records potential difference, compared to a neutral electrode connected to an ear lobe or mastoid
2) bipolar - records potential difference between two paired electrodes
Noise sources EEG
1) muscle activity near active sites, eye movements & blinks
2) eye movement artifacts ⇒ profound effects on frontal brain sites, specifically mid-frontal sites
Which waves of electroencephalography (EEG) are related to emotions in the
brain?
1) right frontal activation - associated with withdrawal stimuli or
negative emotions, e.g., fear or disgust
2) left frontal activation - associated with approach stimuli or positive
emotions, e.g., joy or happiness, ⇒ asymmetrical frontal EEG activity - changes in valence
3) beta bands - related to valence
4) pre-frontal (& parietal) asymmetry in alpha & temporal asymmetry in gamma band - for arousal (&valence) recognition
5) Changes in gamma band are relate to happiness & sadness
6) Decrease in alpha wave in different sides of temporal lobe (left
for sadness, right for happiness)
Gender differences for EEG
1) men rely on recall of past emotional experiences to evaluate current
emotional experiences, have more individual differences among their EEG patterns
2) women, engage emotional system more readily, share more similar EEG patterns among them
Outline the steps involved in emotion recognition from EEG
1) Expose user to test stimulus
2) Record voltage changes in the brain
3) Remove noise & artifacts from recorded signals
4) Analyse resulting data & extract features
5) Train classifier & use computed features ⇒ interpret raw signals
What are the main approaches to removing artefacts from EEG recording?
1) providing
information to participants about their posture
2) manually remove data due to different types of artifacts associated to participants
3) Blind Source Separation (BSS) & Independent Component Analysis (ICA): for removing artefacts due to eye movements, blinks & line noise
4) other methods: Common Average Reference (CAR),
Laplacian, or Average Mean Reference (AMR)
Test protocols for EEG
1) Subjects:
- sufficient number to verify accuracy & meaningfulness of data/results
- balanced number of subjects from each gender
2) Approaches to emotion elicitation:
- subject-elicited, ask participants to remember past emotional episodes of
their life or act as if they were feeling a given emotion
- event-elicited, use different modalities: visual, auditory, tactile or odour stimulation, cover desired arousal levels & valence states
3) influenced by complexity & number of targeted emotions
4) Ground truth of emotional state:
- self-ratings of subjects
use standard stimulus sets,
- duration of an affective phenomenon ranges from full blown emotions (e.g. minutes) to traits, lasting for years if not a lifetime
Sampling freqs for EEG
512, 256, 1024, 2048, 128, 500 Hz
What considerations should be made when selecting EEG features and electrodes?
1) assumptions:
- spectral power in various frequency bands is often
associated to emotional states
- coherence & phase synchronization of pairs of electrodes
- frontal asymmetry in a band power as differentiator of valence levels
2) Most appropriate features not agreed upon
3) features in time, frequency, time-freq domain
4) calculated from recorded signal of single electrode
What does normalised 1st difference feature measure?
captures self-similarities of EEG signal
What do Hjorth features measure?
activity, mobility, complexity
What does non-stationary index (NSI) measure?
complexity by analysing variation of local average over time
What does fractal dimension measure?
measures complexity, produces results closer to theoretical FD than other
methods
What do high order crossing features measure?
1) captures oscillatory pattern of EEG signal
2) applies a sequence of high-pass filters to zero-mean time series Z(t)
Which frequency bands are used to extract power features?
1) 1-4 Hz, delta
2) 4-8, theta
3) 8-10, slow alpha
4) 8-12, alpha
5) 12-30, beta
6) 30-64, gamma
What features can be extracted from the discrete wavelet transform of an EEG
signal?
1) decomposes signal in different approximation & detail levels corresponding to different frequency ranges,
conserves time information of signal
2) ⇐ correspondence of frequency bands & wavelet
decomposition levels depends on sampling frequency
3) features: energy corresponding to the bands
frequencies of α, β and γ, via root mean square
(RMS), recursive energy efficiency
How can features from combinations of electrodes be measured?
1) Magnitude Squared Coherence Estimate (MSCE)
2) hemispherical asymmetry:
- differential, differences in power bands of pairs of electrodes
- rational, use spectrogram to compute power ratios for consecutive time windows
Two types of feature selection methods
1) wrapper methods - iteratively select features based on classifier performance, computationally expensive
2) Filter methods - use measure of info (e.g. variance), select features with variance above threshold, less computational power but ignores cross-feature relationships
Describe the process of feature selection in ReliefF
1) univariate technique
2) uses subsample of instances to adjust weights of each feature depending on ability to discriminate between two classes
3) estimates weight quality for each feature based on difference to nearest hit/miss
4) to account for multiple classes, search n nearest instances from each class weighted by prior probabilitites
Discuss the most commonly used features in emotion recognition from EEG
signals
1) Commonalities/differences between subjects - not very
apparent
2) Most frequently selected feature group is HOC features
3) Rational asymmetry features - most valuable feature type
among the combinations of electrodes
4) Fractal, NSI, Hjorth - measuring complexity
5) Features computed from frequency bands β & γ are
successfully selected more often than other bands (lower fs)
6) Prevalent frequency band powers & equal sized bins show relatively low scores on average
7) Group of statistical features affords a differentiation
between feature types
Discuss the use of sensor technologies to assess neurophysiological activity for
affective computing
1) Measures cortical electric or magnetic fields directly resulting from nerve impulse of groups of neurones (e.g. EEG)
2) Measures metabolic activity within cortical structures (e.g. blood oxygenation, fMRI)
3) EEG:
- relatively unobtrusive
- can be recorded using wearable devices ⇒ increasing mobility & options for locations to collect data
- affordable for private households
- easy to set up
4) Comparable wearable sensor modalities:
- e.g., functional near-infrared spectroscopy (fNIRS)
- not affordable
- not high spatial resolution
Scherer’s component process model
1) Existence of several components of affective responses that reside in central nervous systems (CNS), including:
- processes of emotional event perception & evaluation
- self-monitoring
- action planning & execution
2) ⇒ Brain can differentiate affective states in terms of their neurophysiological characteristics:
- neural responses that occur after encountering an
emotionally salient (important) stimulus
- occur within tens of ms
- not under the volitional control of a person ⇒ reliable in terms to their true nature
- in contrast to the slower physiological responses in the range of seconds & are amenable to conscious influence
The human brain has been considered to comprise two functional circuits. What do they represent?
1) two functional circuits represent:
- sensory information on stimulus event
- somatovisceral impact (influence of the body framework on deep inward
feeling) as remembered/predicted from previous experience
First functional circuit:
1) comprises the following structures for gathering & binding of information from external & internal sensory sources:
- basolateral complex of amygdala
- ventral & lateral aspects of OFC
- anterior insula
2) enabling information exchange about perceived events & objects
- amygdala codes original value of the stimulus
- OFC creates a flexible experience & context-dependent representation of the object’s value
3) insula:
- processes information from inner organs & skin
- forms awareness about state of the body
4) integration of sensory information & information about the body’s state creates a value-based representation of the
event or object
Second functional circuit:
1) VMPFC (including anterior cingulate cortex (ACC) &
amygdala):
- links sensory information about an event, as integrated by first circuit, to its visceromotor outcomes
- ⇒ informs judgements & choices, and is active during
decisions based on intuitions & feelings
Both the first and second functional circuit:
1) project directly & indirectly to hypothalamus & brainstem
2) fast & efficient computation of object values
3) influence autonomous chemical & behavioural responses
What are event-related potentials? What are their characteristics?
1) Event-related potentials (ERPs) - prototypical deflections of recorded EEG trace in response to a specific stimulus event
2) computed by averaging traces following multiple stimulation events of same condition, which reduces sporadic parts of the EEG
trace not associated with the functional processes involved in response to the stimulus, e.g.:
- early potentials, initial perception and auto evaluation of presented stimuli
- late event-related potentials:
- higher level
- more concious
- late positive potential (LPP), associated with stronger perceptive evaluation
Discuss the differences between evoked frequency response and induced frequency responses.
1) Evoked frequency responses
- computed by frequency transformation applied to averaged EEG trace
- ⇒ frequency-domain representation of ERP components
2) Induced frequency responses:
- computed by applying frequency transformation on single EEG traces before averaging the frequency response
- capture oscillatory characteristics of EEG traces
3) In an everyday context:
- use of evoked oscillatory responses is limited
- induced oscillatory responses are for detecting affect based on a single & unique emotional event or period
What are the oscillatory characteristics of EEG activities in delta frequency band (0.5 - 4 Hz)?
1) a correlate of the P300 potential ⇒ detects emotionally salient stimuli
2) increased delta band power in response to more arousing stimuli
What are the oscillatory characteristics of EEG activities in theta frequency band (4 - 8 Hz)?
1) emotionally arousing stimuli increases theta band power over frontal &
parietal regions
What are the oscillatory characteristics of EEG activities in alpha frequency band (8 - 13 Hz)?
1) associated with relaxed & wakeful state of mind
2) increase in alpha power during states of relaxation
What are the oscillatory characteristics of EEG activities in beta frequency band (13 - 30 Hz)?
1) increased beta band activity over temporal regions for positive emotions
2) decrease in beta band power in response to stimuli that had an emotional impact on subjective experience
power increases during the tension of (scalp) muscles that occur during
frowning & smiling
What are the oscillatory characteristics of EEG activities in gamma frequency band (> 30 Hz)?
- increases with increasingly positive valence
- increase in gamma band power in response to high arousing visual stimuli
- increase in gamma activity in response to painful stimuli
- power increases during the tension of (scalp) muscles that occur during
frowning & smiling
What are the main features to be extracted in an affective brain-computer interface?
1) Time-domain features:
- amplitude of stimulus-evoked potentials after a stimulus event, e.g., P300
2) Frequency-domain features:
- frequency band power, e.g., alpha band (8 -13 Hz)
- power in frequency bands < 13 Hz - correlated with
amplitude of event-related potentials
3) Make use of a wide spectrum of frequency bands
What are translation algorithms?
1) Translate selected signal features into a command for the application, e.g. a cursor movement or creation of an emotional label
2) Use machine learning approaches trained to find a
mapping between a number of signal features & emotion
labels
3) Classifiers (e.g., linear discriminate analysis & SVM) needed to adapt to:
- particular user’ signal characteristics
- changes over time and changing context of interaction
- changes in brain activity due to user’s efforts in learning and adapting to the system
BCI operating protocol
1) Guides the operation, e.g.,
switching it on & off if the actions are triggered by the
system, or by the user,
when & in which manner feedback is given to user
2) Other characteristics:
- Whether information is actively produced by user or
passively read by the system
- Whether information is gathered dependent of a specify stimuli event (stimulus dependent/independent)
BCI signal acquisition
1) Invasive measures:
- implanted electrodes or electrode grids
- enable direct recording of neurophysiological activity from cortex
- ⇒ better signal-to-noise ratio
- for patient population
2) Non-invasive measures:
- e.g., EEG, fNIRS, or fMRI
- for healthy population
- easy-to-handle & affordable headsets
