done with section 1

app.py

@@ -7,6 +7,7 @@ _ID_CLUSTER_SCREEN_SHOTS = {
     2: ("cluster_2_of_24_latinx_woman.JPG", "Cluster 2 of 24"),
     18: ("cluster_18_of_24_hispanic_nonbinary.JPG", "Cluster 18 of 24"),
     0: ("cluster_0_of_24_black_woman.JPG", "Cluster 0 of 24"),
+    5: ("cluster_5_of_24_white_unmarked_latinx_man.JPG", "Cluster 5 of 24"),
     6: ("cluster_6_of_24_black_man.JPG", "Cluster 6 of 24"),
     7: ("cluster_7_of_24_pacific_indigenous_man_nonbinary.JPG", "Cluster 7 of 24"),
     3: (
@@ -151,7 +152,8 @@ with gr.Blocks() as demo:
                     This tells us that images that are similar to the ones in this cluster will **likely look like** Hispanic women to viewers.
                     You can cycle through [a few other examples right](https://hf.co/spaces/society-ethics/DiffusionFaceClustering "or even better, visualize them in the app"),
                     such as cluster 19 which mostly features the words *Caucasian* and *man*, different gender term distributions for *African American* in 0 and 6,
-                    as well as clusters like 7 that showcase the limitations of mapping visual features to ethnicity by grouping together *Pacific Islander*, *Indigenous American*, and *Latino*.
+                    as well as clusters like 7 that showcase the limitations of mapping visual features to ethnicity categories
+                    by grouping together *Pacific Islander*, *Indigenous American*, and *Latino*.
                     """
                 )
             with gr.Column(scale=1):
@@ -167,7 +169,7 @@ with gr.Blocks() as demo:
         with gr.Row():
             with gr.Column(scale=1):
                 id_cl_id_2 = gr.Dropdown(
-                    choices=[3, 8, 23, 12, 13, 7],
+                    choices=[3, 8, 23, 12, 13],
                     value=3,
                     show_label=False,
                 )
@@ -181,17 +183,23 @@ with gr.Blocks() as demo:
                 gr.Markdown(
                     """
                     #### [Stereotypical Representations and Associations](https://hf.co/spaces/society-ethics/DiffusionFaceClustering "Select cluster to visualize to the left or go straight to the interactive demo")
-                    - Native american:
-                        - stereotypical (man 3) vs modern and salient (8) vs less stereotypical (23 woman + nonbinary)
-                        - shows the importance of flexibe categories!!!
-                    - Non-binary: stereotype "depends on ethnicity" - associated with only "woman" + haircut + glasses for caucasian, more diverse for others
-                        - different stereotype and power dynamics, but still all the same haircut and glasses (down to the collar!) in cluster 12 - also only associated with women + white
-                        - compare to clusters 13 (black, also +woman), 8 (native american, gender diversity!). Other clusters with NB mostly + visually diverse, + women, except 7 +men
-
-                    You can see that the models reflect many societal biases -- for instance representing Native Americans wearing traditional headdresses,
-                    non-binary people with stereotypical haircuts and glasses, and East Asian men with features that amplify ethnic stereotypes.
-                
-                    This is problematic because it reinforces existing cultural stereotypes and fails to represent the diversity that is present in all identity groups.
+                    
+                    Even before we start leveraging these clusters to analyze system behaviors in other settings,
+                    they already provide useful insights about some of the bias dynamics in the system.
+                    In particular, they help us understand which groups are more susceptible to representation harms when using these systems,
+                    especially when the models are more likely to associate them with **stereotypical representations**.
+                    
+                    Let's look at **clusters 3, 8, and 23** in the 24-cluster setting, as they all primarily feature images whose prompt contains a Native American identity term.
+                    Cluster 3 is made up entirely of images of individuals wearing large traditional headdresses.
+                    It is also the 4th largest cluster across all generations - which suggests that this one feature has a disproportionate importance in representations of these groups.
+                    This behavior further brings to light well known representation issues in the models' training data.
+                    It also showcases the necessity of modeling the diversity of how people map visual features to inferred social characteristics,
+                    as the images in 8 and 23 are more likely to bear some similarity to depictions of people in contemporary professional settings.
+                    
+                    With a different focus, **clusters 12, and 13** help us explore stereotypical representations of non-binary identities.
+                    While both are made up of over 80% images for which the prompt mentions the word *non-binary*,
+                    the similarities in cluster 12 are particularly flagrant upon visual inspection; with all depictions featuring similar glasses and similar haircuts -
+                    neither of which should be a strong indicator of gender.
                     """
                 )
         with gr.Row():
@@ -199,14 +207,22 @@ with gr.Blocks() as demo:
                 gr.Markdown(
                     """
                     #### [Specification, Markedness, and Bias](https://hf.co/spaces/society-ethics/DiffusionFaceClustering "Select cluster to visualize to the right or go straight to the interactive demo")
-                    - Cluster 19: both axes
-                    - Unmarked gender across ethnicity: 6 and 0 have the most AfrAm, 36% vs 18%
-                    - Unmarked ethnicity across genders: 15 has the most unmarked ethnicity of woman>man clusters
+                    
+                    The last phenomenon we study through the lens of our clustering is that of **markedness**,
+                    or *default behavior* of the models when neither gender nor ethnicity is specified.
+                    This corresponds to asking the question: "If I **don't** explicitly tell my model to generate a person of specific genders or ethnicities,
+                    will I still see diverse outcomes?"
+                    
+                    Unsurprisingly, we find that not to be the case.
+                    The clusters with the most examples of both prompts with unspecified gender and ethnicity terms are **clusters 5 and 19**,
+                    and both are also strongly associated with the words *man*, *White*, and *Causian*.
+                    This association holds across genders (as showcased by **cluster 15**, which has a majority of *woman* and *White* prompts)
+                    and ethnicities (comparing the proportions of unspecified genders in **clusters 0 and 6**)
                     """
                 )
             with gr.Column(scale=1):
                 id_cl_id_3 = gr.Dropdown(
-                    choices=[19, 0, 6, 15],
+                    choices=[5, 19, 0, 6, 15],
                     value=6,
                     show_label=False,
                 )