Week 3 Flashcards
Why has camera-based facial expression detection been popular for affective computing?
1) non-intrusive
2) reasonable accuracy
3) inexpensive hardware
How do the five dimensions of cognitive process (Roseman et. al 1990) aid affective computing?
1) Consistency motives - how well an affect-inducing event helps or
hinders one’s intentions
2) Probability - the certainty that an event will actually occur
3) Agency - the entity that produces or is responsible for the event
4) Motivational state - “appetitive” (an event leading to a reward) or
“aversive” (one leading to punishment)
5) Power - whether the subject is (or feels) in control of the situation
Emotions as social constructs…
1) Emotions cannot be explained solely from physiological or
cognitive terms
2) Emotions are primarily social constructs
⇒ social level of analysis
3) Relationship between emotion & language
⇒ universality of Ekman’s studies?
4) Some cultures have emotional labels that cannot be literally translated to English
What does affective neuroscience offer?
1) techniques to understand emotional processes & their neural correlates
2) Offers an understanding of neural circuitry that underlies emotional experience, etiology of certain mental health pathologies
3) Neural substrates of cognition & emotion overlap substantially
4) Emotional learning can occur without awareness
5) Emotional behaviour do not require explicit processing
6) An alternative to self-reports or facial expressions: EEG
How does J.A. Russel (2013) emotion framework bridges the different emotion
theories?
1) core affect - neurophysiological state described as a point in
valence & arousal space
2) importance of context & separating emotional episodes (anger) from emotion categories (anger class)
3) emotions consist of loosely coupled components, e.g.,
physiological responses, bodily expressions, appraisals,
etc.
4) unite categorical models to define
emotions with dimensional models
What challenges are encountered with using Facial Action Coding System?
1) humans having to decompose an expression into set of AUs
2) coding system was originally created for static pictures not auto detection
3) performance of systems that automatically detect the action units do not match humans
What are the concerns with current automated face-based affect detection systems?
1) detect 6 basic emotions
2) most relied on data sets with posed facial expressions
3) few could operate in real time
4) most required pre-segmented emotion expressions
5) almost none integrated contextual cues with facial feature tracking
What are the characteristics of current systems for vocal communication of emotion?
1) pitch as an index to arousal
2) detection accuracy rates lower than facial expressions
3) sadness, anger & fear are best recognised through voice; disgust is the worst
4) some ambiguity w.r.t. how different acoustic features communicate the different emotions
5) most systems trained on spontaneous speech
6) some efforts aimed at detecting other states (i.e., non-basic emotions), e.g., frustration
7) context is often overlooked
8) Voice is a promising signal in AC applications: low-cost, non-intrusive, & fast time resolution
What have body language and posture to offer to affective computing?
1) Human body - large & multiple degrees of freedom ⇒ bare unique configurations
2) Static posture positions can be combined & temporally
aligned with movements
3) Affective state can be decoded over long distances
4) Gross body motions are unconscious & unintentional
⇒ not susceptible to social editing
What are the three approaches to fuse signals from different sensors in a multimodal affective computing system?
1) Data Fusion:
- on raw data for each signal with same temporal resolution
- not commonly used due to sensitivity to noise
2) Feature Fusion:
- on set of features extracted from each signal
- features are individually computed then combined across sensors
3) Decision Fusion
- merges classifier output for global view across sensors
Why is it necessary to incorporate machine learning in affective computing?
1) Some tasks cannot be defined well except by example
2) Hidden in data are important relationships & correlations
3) System design does not work well in the environments in which it is used , on the job improvement
4) Amount of knowledge about certain tasks too large for explicit encoding by human
5) Environments change over time
6) New knowledge about tasks is constantly being discovered
State how a k-NN classifier operate. State its advantage to nearest neighbour
classifier and its limitation
1) k - number of neighbouring samples’ labels to consider
2) Assign sample to the most frequent assigned label
3) Robust to noise
4) Computationally expensive if the dataset is large
State the performance measures commonly used to quantify and compare the characteristics of different estimators. State also the aim of an optimal estimator in terms of these
measures.
1) Expected value of estimate: E[θˆ]
2) Bias of estimate: E[θˆ − θ] = E[θˆ] − θ
3) Covariance of estimate: Cov[θˆ] =
E[(θˆ − E[θˆ])(θˆ − E[θˆ])’]
4) Optimal estimators aim for zero bias & minimum estimation error covariance
What is meant by prior space and posterior parameter space?
1) Prior space - collection of all possible values that the parameter vector can assume
2) Posterior parameter space - subspace of all likely values of a parameter consistent with both prior information & evidence in the observation
