Sensing Valence and Confusion with Facial EMG Phil Davis and Hsin-Ni Ho Outline I. II. Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Outline I. II. Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Experimental Setup Electrodes attached to: Corrugator supercilii (“brow”) Zygomaticus major (“cheek”) Grounded on arm Outline I. II. Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Outline I. II. Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Problems Addressed Can we sense valence from facial electromyographic (EMG) activity? Intensity? Can we sense valence from mouse behavior (pressure in particular)? In general, what types of activity can we sense with facial EMG? Relevant Prior Work Recent study by Larsen, Norris, Cacioppo Change in mean “cheek” EMG activity correlated positively with valence Change in mean “brow” EMG activity negatively correlated with valence (CS = “brow” muscle, ZM = “cheek” muscle) Outline I. II. Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Experimental Procedure 6 subjects General procedure: 1. Subject watches a short film clip 2. Subject answers questions about film content and experienced affective state Subject uses pressure sensitive mouse EMG output is recorded during all stages Film Clip Sequence Each film clip was roughly 3 minutes 1. Golf Instruction (Neutral) 2. Clip from “The Champ” (Negative) Sadness in 94.2% (Gross & Levenson) 3. Golf Instruction (Neutral) 4. Clip from “Robin Williams Live” (Pos) Amusement in 84.1% (Gross & Levenson) Did not vary order Outline I. II. Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Self-reported Valence Consistent with Expectations Neutral Valence: 4 subjects indicated “neutral” affect, 2 subjects indicated low intensity positive affect Negative Valence: All 6 reported sadness Positive Valence: 5 of 6 reported amusement The sixth subject did not like Robin Williams Example of Output Output normalized due to large scale differences Mean Output Consistent with Prior Work Individual Results Varied Other Observations For some subjects, brow activity was a good indicator of web form activity Outline I. II. Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Discussion Highly sensitive to individual subject differences and electrode positioning Mean output consistent with Larsen, Norris, Cacioppo Do results extend to other positive/negative states? Unable to correlate EMG output with self-reported mood intensity EMG output may be useful as an activity recognition sensor Unable to correlate mouse pressure or velocity with valence Dynamic model of mouse behavior may produce better results? Outline I. II. Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Outline I. II. Introduction and Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Problem Addressed Does the facial expression ‘frowning’ represent the feeling of confusion? Can we recognize the feeling of confusion with the EMG outputs? Does the importance of understanding influence the feeling of confusion? Outline I. II. Introduction and Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Experimental Procedure 6 subjects General procedure: 1. Subject listens to an audio clip 2. Subject answers questions about audio content and experienced affective state EMG output is recorded during all stages Eliciting Confusion with Audio Recordings Two levels of the importance of understanding: Low level importance of understanding. High level importance of understanding. Audio clips to induce different levels of confusion: A B Mood control 40 s classic music Control spoken with standard American English Confused by accent spoken with accent Confused by meaning A brief paragraph with confusing content spoken with American English Ratings and Measurements for feeling of Confusion Subjective rating 5 scale Self-report confusion rating Objective rating Test of understanding Measurement EMG responses Outline I. II. Introduction and Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Result: EMG responses Brow Cheek Mean 210.14 2654 Std 214.44 5876 Difference in importance of understanding NO DF=1, P=0.28 NO DF=1, P=0.24 Result: Average EMG responses For brow Brow More EMG activity when 0.8 0.7 Normalized EMG 0.6 mood control 1L 0.5 1W 0.4 listening to clips with accent 2L 2W 0.3 3L EMG in ‘web-filling’ parts increase with level of confusion in high importance of understanding 3W 0.2 0.1 0 important not important Cheek For cheek 0.8 0.7 Normalized EMG 0.6 0 1L 0.5 1W 0.4 2L 2W 0.3 3L 3W 0.2 0.1 0 important filling the web form not important EMG activity increase with degree of confusion ->Subject started to laugh Result: Self confusion rating and test score Average plot Rating and Score Confusion and Test Score 6.00 5.00 4.00 3.00 2.00 1.00 0.00 Confusion Test 1 2 3 4 Audio section 5 6 Self rating confusion Test score Mean 3.81 2.38 Std 1.51 1.87 Difference for 6 clips difference among subjects create the feeling of confusion consistently Yes DF=5, P < 0.0001 NO DF=5, P=0.42 Yes DF=5, P < 0.0001 NO DF=5, P=0.33 Outline I. II. Introduction and Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Discussion Huge individual difference in EMG responses -> normalize the data More Brow activity when filling web forms -> It is more correlated to ‘the process of thinking about confusion thing’ More Brow activity when listening to clip with accent Cheek activity increase with level of confusion -> So confused that subjects were giving up Difference in importance of understanding Outline I. II. Introduction and Experimental Setup Valence A. Problems Addressed B. Experimental Procedure C. Results D. Discussion III. Confusion A. Problems Addressed B. Experimental Procedure C. Results D. Discussion IV. Conclusions Conclusions Baseline difference in EMG for different muscles Individual difference in EMG response EMG response sensitive to electrode positioning Brow EMG is negatively correlated with valence and positively correlated with feeling of confusion Cheek EMG is positively correlated with valence Facial EMG may be useful for activity recognition