Academia.eduAcademia.edu

Figure 3 – uploaded by Amy Alexander

See full PDFdownloadDownload figure
The artworks in Deep Hysteria redeploy the bias embedded in facial analysis algorithms in the service of probing this deeply entrenched social bias. Fig. 3. Deep Hysteria artwork

Figure 3 The artworks in Deep Hysteria redeploy the bias embedded in facial analysis algorithms in the service of probing this deeply entrenched social bias. Fig. 3. Deep Hysteria artwork

Related Figures (11)

Hysteria, Then and Now Fig. 1. Sequence of drawings from 1893 depicting a woman with “hysteria.”
Hysteria, Then and Now Fig. 1. Sequence of drawings from 1893 depicting a woman with “hysteria.”
Fig. 2. Google Search Image Results for “Anxiety,” January 2023 The stereotype is further augmented by the cultural expectation of smiling as women’s default facial expression. Consider the phenomena of “Resting Bitch Face” (Grossman 2019) and “telling women tc smile.” (Smith 2016) A neutral facial expression on a woman is read as disgust, distress or unhappiness: “What’s wrong?”
Fig. 2. Google Search Image Results for “Anxiety,” January 2023 The stereotype is further augmented by the cultural expectation of smiling as women’s default facial expression. Consider the phenomena of “Resting Bitch Face” (Grossman 2019) and “telling women tc smile.” (Smith 2016) A neutral facial expression on a woman is read as disgust, distress or unhappiness: “What’s wrong?”
Fig. 4. Deep Hysteria artwork In addition to their utilitarian function of portraying gender-variable artificial personae, Deep Hysteria’s virtual portraits also serve as an exploration of texture, pose, and identity in generative deep learning portraiture. While many portraits generated via deep learning default to the use of posed, high resolution still photographs as training data, Deep Hysteria is trained on frames from You Tube vlogs produced by video makers with low subscriber counts. These vloggers, who come from countries around the world, photograph themselves speaking to the camera and are primarily amateur videographers. Their videos are typically shot with phone cameras and webcams and are photographed in a variety of settings, often with less-than-ideal lighting. Consequently, Deep Hysteria’s generated images are composite portraits of the self-selected vlogger addressing their audience, rather than the posed and curated subject of a photographer’s gaze.”
Fig. 4. Deep Hysteria artwork In addition to their utilitarian function of portraying gender-variable artificial personae, Deep Hysteria’s virtual portraits also serve as an exploration of texture, pose, and identity in generative deep learning portraiture. While many portraits generated via deep learning default to the use of posed, high resolution still photographs as training data, Deep Hysteria is trained on frames from You Tube vlogs produced by video makers with low subscriber counts. These vloggers, who come from countries around the world, photograph themselves speaking to the camera and are primarily amateur videographers. Their videos are typically shot with phone cameras and webcams and are photographed in a variety of settings, often with less-than-ideal lighting. Consequently, Deep Hysteria’s generated images are composite portraits of the self-selected vlogger addressing their audience, rather than the posed and curated subject of a photographer’s gaze.”
Fig. 5. Sequence of progressively gender-varied Deep Hysteria raw images with Amazon Rekognition- generated emotion/gender/age labels While the predominant mood analyzed by Rekognition for both male-identified and female-identified “neutral expression” images was “calm,” a significantly greater number of male-identified images received this designation. Female-identified images were more likely to be analyzed with stereotypically female emotions: “fear,” “confused,” “sad,” “surprised,” “disgusted.” Male-identified images were more likely to be designated “angry” by Rekognition. Although Deep Hysteria faces are gendered across the gender spectrum and thus includes non-binary faces, Rekognition lacks designations beyond binary gender presentation; it labels all images as either “male” or “female.” Non-binary-appearing images inclusive, those images Rekognition identified as “female” were more likely to be labeled with stereotypically feminine emotions. This effect was notable in cases where the corresponding masculine image was labeled “calm.”
Fig. 5. Sequence of progressively gender-varied Deep Hysteria raw images with Amazon Rekognition- generated emotion/gender/age labels While the predominant mood analyzed by Rekognition for both male-identified and female-identified “neutral expression” images was “calm,” a significantly greater number of male-identified images received this designation. Female-identified images were more likely to be analyzed with stereotypically female emotions: “fear,” “confused,” “sad,” “surprised,” “disgusted.” Male-identified images were more likely to be designated “angry” by Rekognition. Although Deep Hysteria faces are gendered across the gender spectrum and thus includes non-binary faces, Rekognition lacks designations beyond binary gender presentation; it labels all images as either “male” or “female.” Non-binary-appearing images inclusive, those images Rekognition identified as “female” were more likely to be labeled with stereotypically feminine emotions. This effect was notable in cases where the corresponding masculine image was labeled “calm.”
Fig. 6. Deep Hysteria generated artwork
Fig. 6. Deep Hysteria generated artwork
Fig. 7. Deep Hysteria generated artwork
Fig. 7. Deep Hysteria generated artwork
Table 1. Rekognition analysis of generated neutral expression faces — male Table 2. Rekognition analysis of generated neutral expression faces — female
Table 1. Rekognition analysis of generated neutral expression faces — male Table 2. Rekognition analysis of generated neutral expression faces — female
Fig. 8. Deep Hysteria generated artwork Rough statistics across the broader dataset The statistics listed below are based on an informally collected/generated dataset. The sample size is small, and the generation and collection methodologies are not designed with the necessary rigor for scientific research. In particular, the broad gender spectrum of generated images sometimes exhibits apparent racial shifts and other anomalies at the extremes that could impact accuracy in some cases. Overall, however, they do give a general indication of the bias observed in the development of the Deep Hysteria artworks.
Fig. 8. Deep Hysteria generated artwork Rough statistics across the broader dataset The statistics listed below are based on an informally collected/generated dataset. The sample size is small, and the generation and collection methodologies are not designed with the necessary rigor for scientific research. In particular, the broad gender spectrum of generated images sometimes exhibits apparent racial shifts and other anomalies at the extremes that could impact accuracy in some cases. Overall, however, they do give a general indication of the bias observed in the development of the Deep Hysteria artworks.
Smile! The original stills of actual YouTube video makers used in the training dataset were generated were also analyzed by Rekognition over the course of a year, as part of the What the Robot Saw (Alexander 2020) live stream artwork. These images were submitted to Rekognition as raw images without algorithmic variations. Unlike the generated images analyzed, which were limited to neutral expressions, these images incorporate the actual range of vlogger facial expressions. The following statistics summarize the apparent emotions Rekognition identified in the actual vloggers. Notable in these statistics: female vloggers were much more likely to be labeled as “happy.” An informal visual analysis by the artist found that most images labeled as “happy” did indeed appear to be presenting as “happy” — i.e. more women smiled in their videos. This observation appears to support the assumption that women perceive more social expectation than men to smile in their public online presentation.
Smile! The original stills of actual YouTube video makers used in the training dataset were generated were also analyzed by Rekognition over the course of a year, as part of the What the Robot Saw (Alexander 2020) live stream artwork. These images were submitted to Rekognition as raw images without algorithmic variations. Unlike the generated images analyzed, which were limited to neutral expressions, these images incorporate the actual range of vlogger facial expressions. The following statistics summarize the apparent emotions Rekognition identified in the actual vloggers. Notable in these statistics: female vloggers were much more likely to be labeled as “happy.” An informal visual analysis by the artist found that most images labeled as “happy” did indeed appear to be presenting as “happy” — i.e. more women smiled in their videos. This observation appears to support the assumption that women perceive more social expectation than men to smile in their public online presentation.
Table 3. Rekognition analysis of actual YouTuber faces (any expression) —male
Table 3. Rekognition analysis of actual YouTuber faces (any expression) —male
Table 4. Rekognition analysis of actual YouTuber faces (any expression) —femal
Table 4. Rekognition analysis of actual YouTuber faces (any expression) —femal
Related topics:
Digital ArtsComputational ArtAlgorithmic BiasGenerative AI
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved