{"id":25002,"date":"2026-05-04T11:30:57","date_gmt":"2026-05-04T16:30:57","guid":{"rendered":"https:\/\/hexagram.ca\/demo71-david-drouin\/"},"modified":"2026-05-19T15:02:50","modified_gmt":"2026-05-19T20:02:50","slug":"demo71-david-drouin","status":"publish","type":"post","link":"https:\/\/hexagram.ca\/en\/demo71-david-drouin\/","title":{"rendered":"DEMO71 David Drouin \u2013 R\u00e9sonance latente"},"content":{"rendered":"\n

May 2026<\/strong><\/p>\n\n\n\n

<\/div>\n\n\n\n

R\u00e9sonance Latente<\/em> is a series of interactive audiovisual (A\/V) improvisational stage performances between David Drouin<\/a>\u2014electronic musician and student member\u2014and an artificial intelligence agent responsible for the visual counterpart. The performances took place in the context of a research\u2011creation project exploring the potential of machine\u2011learning techniques to augment A\/V improvisational capacities. This project was conducted as part of a master\u2019s degree in Communication (research\u2011creation in experimental media profile) at Universit\u00e9 du Qu\u00e9bec \u00e0 Montr\u00e9al.<\/p>\n\n\n\n

The project examines how engaging with these techniques can benefit from their capacity for emergent novelty while maintaining modalities of control that foster sensemaking. In this DEMO, David Drouin outlines the trajectory of this research.<\/p>\n\n\n\n

<\/div>\n\n\n\n
\n\n\n\n
<\/div>\n\n\n\n

Improvisational Fictionalism and the Aesthetics of Behavior<\/strong><\/strong><\/h2>\n\n\n\n
<\/div>\n\n\n\n

Philosopher Eric Lewis<\/a> offers an interesting perspective for addressing the question of sense in musical improvisation with a machine through his theory of improvisational fictionalism<\/em> (Lewis, 2019). He argues that there is no loss of authenticity in improvising with a machine if it possesses the capacity to produce convincing results. According to Lewis, one can therefore \u201ctreat them as imaginary partners in our improvisations, playing a game of make-believe.\u201d<\/p>\n\n\n\n

To determine what might constitute convincing results in the context of a visual agent, I draw on the aesthetics of behavior<\/em> (Penny, 2000). This framework emphasizes that designing behaviors involves creating a contingent model of what might happen in the world and how a system can respond to it in order to guide the user\u2019s aesthetic attention in a direction coherent with the work itself. This leads to an aesthetic negotiation between interaction dynamics and authorial intentions. This perspective makes it possible to analyze the degrees and morphology of the deployment of an agent\u2019s behaviors over time (Audry, 2021). This grounding guided my creative choices toward setting up conditions that might encourage the agent\u2019s emergence of behaviors capable of adopting recognizable forms and metamorphosing over time. Ultimately, it is through my observation of the machine\u2019s behaviors within the sociocultural context of performance that I am able to evaluate its situated production of novelty<\/em> (Bown, 2021) and judge whether these form convincing results “that open the game of make\u2011believe<\/em>.”<\/p>\n\n\n\n

<\/div>\n\n\n\n

The visual agent<\/strong><\/h2>\n\n\n\n
<\/div>\n\n\n\n

The agent consists of an assemblage of software using artificial neural networks that I programmed and trained to produce figurative video collages based on the analysis of human musical gestures. These collages draw on a corpus of video archives from Beauce (DeLessard, n.d.), my region of origin. This material was selected for its transbiographical<\/em> qualities\u2014embedding the biographical within the fictional\u2014as well as for its strong potential to generate poetic meaning during improvisation.<\/p>\n\n\n\n

<\/div>\n\n\n\n
\"Sch\u00e9ma
Operating diagram of the visual agent. Photo : \u24b8David Drouin.<\/sup><\/figcaption><\/figure>\n\n\n\n
<\/div>\n\n\n\n

The training process involves associating musical analysis\u2014such as sound spectrum or rhythmic analysis of an instrument\u2014with visual\u2011effect parameters, such as the repetition rate of a video loop or the selection of an image to display. These associations are generated through shared descriptors that I manually assigned, or by allowing neural\u2011network algorithms to classify data based on patterns they determine statistically. These trainings were conducted using small datasets<\/em> (Abuzuraiq & Pasquier, 2024) that I created or assembled myself. This approach facilitates navigation and manipulation of the neural networks throughout the training process and allows me to better grasp the influence of creative decisions on their behavior.<\/p>\n\n\n\n

<\/div>\n\n\n\n
\n
\n
\"Aper\u00e7u<\/figure><\/div><\/div>\n\n\n\n
\n
\"Aper\u00e7u<\/figure><\/div><\/div>\n<\/div>\n\n\n\n
<\/div>\n\n\n\n

Overview of the creation process for training data through qualitative associations. Photos : \u24b8David Drouin<\/sup><\/p>\n\n\n\n

<\/div>\n\n\n\n

To generate a variety of behaviors, a central algorithm orchestrates the musical signals that trigger visual events. It also determines how these signals can be analyzed through trained neural networks and properties such as cadence, delaying, or smoothing of each analysis. In the event of silence from a signal it has selected, the algorithm falls back on active musical signals in order to create a convincing machine listening effect of the ongoing musical performance.<\/p>\n\n\n\n

<\/div>\n\n\n\n
\n