Despite the ubiquity of cluster analysis, not all data consists of multiple natural clusters, and thus clustering is not always pertinent. Testing for clusterability – the degree to which a dataset includes intrinsic cluster structure – can inform whether clustering techniques should be applied. Clusterability evaluation is complicated by the inherently ambiguous nature of clustering and dependence on the clustering requirements of the underlying application. We present the first comprehensive clusterability package, which implements clusterability tests in R. Diverse clusterability tests enable users to select a technique for their specific needs. Detailed installation instructions, examples and R code are included.

Text-to-image models, like Midjourney, DALL-E, and Stable Diffusion, have been shown to reinforce harmful biases, often perpetuating outdated and discriminatory stereotypes. In this study, we delve into a particular bias largely overlooked in generative image research: Brilliance Bias. By age 6, many children begin to internalize the damaging notion that intellectual brilliance is a male trait—a belief that persists into adulthood. Our findings demonstrate that popular image AI models possess this bias, further entrenching the misguided notion that exceptional intelligence is inherently male. This study calls for addressing brilliance bias in AI to ensure a more realistic representation of intellectual capabilities, helping shape a future where talent and brilliance are more broadly recognized.

The Computational Creativity literature stresses the importance of evaluation in computational creativity systems in addition to generative capabilities. But, what about machines that only evaluate? When it comes to co-creative systems, humans often take on the primary evaluative role, while machines assist with the generation of creative artifacts. In this paper, we propose flipping the paradigm, envisioning machines that (only) evaluate humans in the creative space. We call such machines Creative Evaluators whose end goal is to evaluate human-made artifacts and provides feedback to aid humans in the creative process, while the machine itself refrains from directly generating any content. We present a co-creative framework for designing Creative Evaluators, which includes the tension between originality and quality of creative artifacts, and the need for explainability and fairness in machines that take on an evaluative role for human-made artifacts. To illustrate this framework, we present Titlevator, a machine that evaluates titles for AI papers, while also providing feedback to the user on the originality and the quality of the paper being evaluated.

Language has a profound impact on how we perceive the world. With GPT-3’s rise in popularity, as of latest utilized in 300 applications averaging 4.5 billion words per day, it is critical to identify and correct biases in its generations. A variety of biases have been identified in generative language models, spanning biases based on gender, race, and religion. In this paper, we pioneer the study of the Brilliance Bias for generative models. This implicit, yet powerful bias imposes the idea that “brilliance” is a male trait and in turn, sets back women’s achievements starting as early as ages 5-7. We perform an analysis of two GPT-3 models, the base GPT-3 model (davinci) and InstructGPT (text-davinici-002), focusing on adjectives, verbs and lexicons found in their generations. Our analysis reveals the presence of substantial Brilliance Bias across both models.

Recent years have seen a budding interest in therapeu- tic applications of creative machines, spanning both autonomous systems and agents that enrich the human cre- ative process. This paper takes a deep dive into therapeutic modalities through the lens of computational creativity and explores opportunities in this exciting emerging domain. In addition to bringing to light to how computational creativity can interface with mental health and wellness, the current paper brings atten- tion to the potential risks and pitfalls of bringing creative machines into the therapeutic context. We hope that this work, conducted in collaboration between CC researchers and practising psychotherapists, will help pave the way forward to responsible and effective applications of computational creativity to therapeutic do- mains.

ABSTRACT Self-expression is central to mental well-being and mental health therapy. Art therapy offers a wide range of expressive mechanisms, allowing individuals to process their emotions when traditional therapies prove unsuccessful. However, a lack of expertise or comfort with artistic expression, along with cost and waiting times, may hinder one's ability to receive needed mental health support. Creative machines can offer novel therapeutic approaches enabling the bereaved to engage in creative expression as and when needed. In this paper, we apply a co-creative songwriting system, ALYSIA, as a new form of therapy for those who had recently suffered the loss of a loved one. We evaluate the utility of this creative system in aiding bereaved individuals through user testing. The utility of collaborative creative systems for adaptation to bereavement is discussed and may have implications for other therapeutic applications.

•Properties that help solve the clustering users dilemma.•Weighted properties formally differentiate clustering methods.•A formal classification highlights advantages of center-based clustering techniques.

Self-expression is essential to processing our thoughts and feelings and is central to successful mental health therapy. Art therapy provides a wider range of expressive mechanisms than offered through traditional approaches, allowing individuals to process their emotions when traditional therapies prove unsuccessful. Yet, effective expression through art therapy may call on a level of artistic experience that is not available to all. As such, a lack of expertise or comfort with artistic expression may hinder one’s ability to receive needed mental health support. Creative machines can offer novel therapeutic approaches by offloading the need for creative expertise and opening up creative self-expression to those who lack the corresponding experience. In this paper, we focus on bereavement, and explore a co-creative songwriting system, ALYSIA, as a new form of therapy for those who had recently suffered the loss of a loved one. We evaluate the utility of this creative system in aiding bereaved individuals through several case studies. In addition, we discuss the utility of co-creative systems to the therapeutic context with potential application to a broad range of therapies.

Casual creators offer an enjoyable and readily accessible creative experience by enabling a safe and easy exploration of a creative space. This explorative and intrinsically motivating process lends itself to education applications, giving rise to a new category of casual creators, which we call Educational Creators. We illustrate this concept through EarthMood, an educational creator for deepening students’ understanding of Climate Change. We demonstrate EarthMood through historical data, showing the deterioration of the planet, as well as on recent data, illustrating an improvement in climate due to COVID-19. The remotely-accessible nature of educational creators makes them applicable to both traditional and remote learning settings.

This work proposes a methodology for conducting field work in computational creativity, referring to field work as the effort of actively making a system or its artifacts widely accessible outside the academic world. Field work aims to study how creative computer agents and/or their products influence society, and how the dynamics that arise from the interaction between people and those inventive machines or their artifacts can inform the design of computational creativity methods, systems and artefacts. In this paper, we reflect on our experiences making our systems ALYSIA and MEXICA broadly available. In the case of ALYSIA, the system itself was made accessible, whereas MEXICA’s artifacts (stories) were shared through a traditionally published book for a broad readership. We then propose a five step methodology for effectively conducting field work in Computational Creativity. The participation of the computational creativity community is essential to test and enrich this methodology.

This paper introduces EMILY, a machine that aims to create original poems in the style of renowned poet Emily Dickinson. Dickinson’s succinct and syntactically distinct style with unconventional punctuation makes for an interesting challenge for automated poetry creation. A user study compares EMILY’s poems to Emily Dickinson originals, demonstrating the machine’s ability to evoke mental images and highlighting challenges for future work. Introduction Poetry writing is an artform dating back to prehistoric times (Finnegan 2012). A successful poem elicits imagery and evokes emotion through an interlock of relationships between semantics, syntax, grammar, punctuation, rhythm and rhyme. Machine generated poetry is itself an artform distinct from human made poetry, with computer generated poems created across human languages through a variety of computing techniques (see, for example, (Lau et al. 2018), (Zhang and Lapata 2014) and (Hämäläinen and Alnajjar 2019)). While poetry machines often create original works without focus on any particular poet, there are exceptions. Style imitation has, for example, been applied to the works of Italian poet Dante Alighieri (Zugarini, Melacci, and Maggini 2019), Bob Dylan lyrics (Barbieri et al. 2012), and the works of William Shakespeare and Oscar Wilde, amongst several others (Tikhonov and Yamshchikov 2018). Poetic style imitation offers the opportunity to immortalize a poet by keeping their voice alive through novel works. From an evaluation standpoint, the generated works can be compared with those of the original creator, enabling a variation of the CC Turing Test by checking whether unbiased observes are able to discern generated artifacts from original ones. Other variations involve comparing the original and generated works on important criteria (ex. stylistic elements of poetry) to help identify where improvement is needed. One of the greatest English poets, Emily Dickinson (1830-1886), is known for effectively capturing feeling and imagery using few words (Emily Dickinson Museum 2020). Dickinson’s style is revealed through unique use of punctuation, syntax, formatting and rhyme (Emily Dickinson Museum 2020). Her succinct and potent poetry makes Dickinson an interesting challenge for style imitation. Figure 1: The poet Emily Dickinson (1830-1886). Photo Credit: Yale University Manuscripts Archives Digital Images Database. In this paper, we present EMILY, a poetry machine that aims to replicate the style of Emily Dickinson’s poems. We present the methodology behind EMILY, along with a user study that compares machine-created poems with Emily Dickinson originals on several poetic criteria. Method The making of EMILY consists of data preprocessing, the creation of custom Markov Chains, and postprocessing. These steps are detailed below. Data Preprocessing EMILY was trained on publicly available Emily Dickinson poetry from the Gutenberg project: “Poems by Emily Dickinson, Three Series, Complete by Emily Dickinson” (Dickinson 2004; Project Gutenberg ). The data was made of 444 poems, consisting of 10178 lines. Punctuation meaningfully contributes to Dickinson’s unique style and as such deserves careful treatment. We saved commas, periods, question marks, and semi-colons. Dickinson is well known for her uses of dashes (Emily Dickinson Museum 2020), which were also preserved. Some Proceedings of the 11th International Conference on Computational Creativity (ICCC’20) ISBN: 978-989-54160-2-8 243 punctuation, particularly all brackets, were omitted, as they introduced noise without helping to capture Dickinson’s style. Dickinson used to number instead of title most of her poems. We discarded all roman numerals in our preprocessing since our focus is on generating the poems’ bodies. The final preprocessing step was to convert any fullycapitalized words found in the poem titles into lower case. This helped to enrich the data set of Dickinson’s words. Words that start with capital letters were left unchanged because Dickinson used capitalized words in the middle of sentences (Emily Dickinson Museum 2020). Custom Markov Chains To endow EMILY with Dickinson’s style, we chose to build our own custom Markov Chains. This gave us greater control over the creative process, particularly as it pertains to punctuation, which is a central element of Dickinson’s poetry. (Barbieri et al. 2012) also observed that unmodified Markov Chains were insufficient for capturing style, in their case as it pertains to Bob Dylan’s use of rhyming. The Markov Chains implementation relies on a dictionary. We create the Markov Chains by iterating through all the words and reading them in reverse. Starting with the first word, we iterate for each word at index i checking if the prior word appears in the dictionary. If so, we add the word to its list of values. If the word before it is not in the dictionary, we add it to the dictionary and start its list of values with the current word as the first word. As a result, we map each word to all the words that proceed it in Dickinson’s writing. Doing so lets us capture the relationship of what words show up after each specific word along with their frequency. Words with higher frequency have a higher probability of being generated. Our final dictionary had a total of 8610 keys. Markov Chains are used to generate the sequence of words for the poems. We format the generated words in the postprocessing phase. Starting Word For single stanza poems, we randomly select the initial word from all words used in Dickinson’s writing. If the poem has more than one body, we rely on the final word in the previous body in order to generate the first word in the sequence body using the Markov process. Body Each stanza in a poem is 20 words long. This keeps the poems at approximately the length of Dickinson’s poems, which consist of short stanzas of 4-5 lines each with 5-6 words per line. The number of stanzas generated for each poem is determined by a variable n passed to EMILY. Closing Word To help bring out Dickinson’s style, concluding words were chosen from amongst those that had punctuation. Postprocessing: Formatting the Poems Not only is the choice of words in the poem important to capturing Emily Dickinson’s style, but the format of the poem brings in important stylistic elements. We format the poems based on an analysis of Dickinson’s poetry. Dickinson starts poems with capitalized words, and also follows periods, exclamation marks, or question marks with capitalized word. Words that follow a comma or semi-colon are generally lowercase. More importantly, Dickinson is known for capitalizing words in the middle of sentences, not only words that begin a new line (Emily Dickinson Museum 2020). We traverse through the final list of words and set a flag based on the type of punctuation to determine if the following word should start with a capital or lowercase letter. Following Dickinson’s style (Emily Dickinson Museum 2020), any capitalized words not preceded by a comma or semicolon are left unchanged. The generated list of words is then divided into 5 word sentences, and the first letter of each sentence is capitalized. User Study We evaluate EMILY by comparing its machine-created poems to Emily Dickinson originals on several criteria. This study seeks to gain an initial understanding on the quality of EMILY’s poems. Larger and more in depth studies are left to future work. We surveyed 17 participants, 9 female and 8 male. On a scale of 0-5, 0 being “Not at all Familiar” with Emily Dickinson’s poetry and 5 being “Extremely Familiar”, 3 participants responded with a 4, 5 responded with a 3, 4 with a 2, 1 with a 1 and 4 with a 0. Participants were presented with a total of 12 poems, consisting of 10 of EMILY’s poems and 2 poems by Emily Dickinson. The original poems are Poem 6, “Faith” is a fine invention, and Poem 12, Come Slowly—Eden, which capture many of her stylistic elements. The choice of questions was influenced by previous work evaluating machine-made poetry (Zugarini, Melacci, and Maggini 2019; Hämäläinen and Alnajjar 2019; Lamb, Brown, and Clarke 2015). For each of the 12 poems, participants were asked the following: 1. Is this a typical poem? 2. Is this poem understandable? 3. How much do you like the word choice in the poem? 4. Does the text evoke mental images? 5. Does the text evoke emotion? 6. Do you like this poem? Each question was answered by selecting from a Likert scale: Strongly disagree (0), disagree (1), neutral (2), agree (3), strongly agree (4). The scores of each question were averaged across all respondents for each poem, as shown in Figure 2. The scores of each question were also averaged across all generated poems versus the original Emily Dickinson poems, shown in Figure 3. Results Our survey shows that question 4, “Does the text evoke mental images?”, had the highest average score of 2.17 of all questions for generated poems. Furthermore, the average score of question 4 outranked the average score for Emily Proceedings of the 11th International Conference on Computational Creativity (ICCC’20) ISBN: 978-989-54160-2-8 244 Figure 2: Average scores of questions for each poem based on the Likert Scale. Poems 6 and 12 correspond to original Emily Dickinson poems, while the others were created by EMILY. Dickinson’s poems in 2 of the generated poems. Poem 1, 7, and 10 had the highest score for question 4 as seen in Figure 2. Poems 1, 7 and 10 appear at the end of this section. Three of our generated poems resulted in at least 3 out of the 5 questions averaging to a score higher than 2, in a range similar to Emily Dickinson’s poems’ average scoring of 2-3 (Poem 1, 3, and 11). Each of these poems performed well on a different set of questions. The question “Is this poem understandable?” resulted in the lowest average score across all our generated poems as seen in Figure 3, with a score of 1.43 across all generated poems. Dickinson’s poems averaged to a sco

Numerous scientific disciplines, e.g. social sciences, benefit from field work. What does field work look like in Computational Creativity and what are its potential benefits to research? We refer to the effort of actively making a system or its artifacts widely accessible outside the academia world, and as such getting feedback, as ‘field work’. In this paper, we reflect on our experiences taking our systems, Alysia and MEXICA, out into the wild terrain by making them broadly available. In the case of Alysia, the system itself was made accessible; MEXICA’s artifacts (stories) were shared through a traditionally published book for a broad readership. We consider the utility of field work for these vastly different systems on the CC continuum (Pérez y Pérez 2018), and discuss potential benefits to other research in the area. Finally, we discuss the necessity of developing methodology to enable rigorous registration of knowledge arising from field work in Computational Creativity.

The manuscript describes and visualizes datasets from the datasets package in the R statistical software, focusing on descriptive statistics and visualizations that provide insights into the clusterability of these datasets. These publicly available datasets are contained in the R software system, and can be downloaded at https://www.rproject.org/, with documentation provided at https://stat.ethz.ch/R-manual/R-devel/library/datasets/html/00Index.html. Further information on clusterability is found in the companion to this article, To Cluster or Not to Cluster: An Analysis of Clusterability Methods? (https://doi.org/10.1016/j.patcog.2018.10.026).Brief descriptions and graphs of the variables contained in each dataset are provided in the form of means, extrema, quartiles, standard deviation and standard error. Two-dimensional plots for each pair of variables are provided. Original references to the data sets are included when available. Further, each dataset is reduced to a single dimension by each of two different methods: pairwise distances and principal component analysis. For the latter, only the first component is used. Histograms of the reduced data are included for every dataset using both methods. (c) 2019 The Authors. Published by Elsevier Inc.

Music and art therapy have long been used to treat anxiety disorders, yet their efficacy is limited by a lack of intratreatment monitoring. The integration of EEG monitoring with generative music and art opens the possibility to a new form of more effective and accessible therapies. In this paper, we introduce Myndala, an EEGdriven immersive auditory and VR experience that assists subjects in reaching a relaxed mental state. Myndala was exhibited to broad audiences as part of Night of Ideas “Facing Our Times: The City of The Future” at the San Francisco Public Library. In addition to discussing Myndala and the exhibition, we propose directions for future work that arise from the integration of EEG with generative music and art.

MindMusic explores a new form of creative expression through brain controlled musical improvisation. Using EEG technology and a musical improviser system, Impro-Visor (Keller, 2018), MindMusic engages users in musical improvisation sessions controlled with their brainwaves. Brain-controlled musical improvisation offers a unique blend of mindfulness meditation, EEG biofeedback, and real-time music generation, and stands to assist with stress reduction and widen access to musical creativity.