Threshold Navigator

Timeline

Status: Active

Description

Threshold Navigator is an installation about the instability of classification. At its center, a single fruit rests inside a transparent cube. A camera continuously observes it. A neural network attempts to name what it sees: maracuyá or passion fruit. The fruit does not change. The system’s decision does. Both names refer to the same object. Yet they carry different cultural, geographic, and historical trajectories. “Maracuyá” speaks from Latin American ecologies and vernacular knowledge. “Passion fruit” emerges from colonial translation and missionary taxonomy. The difference is not optical — it is epistemic. What is at stake is not what the fruit is, but how it is situated within language and power.

The installation makes this instability visible and tangible.

The space is illuminated by a reactive light that shifts according to the presence of the viewer. As someone approaches, the light subtly changes intensity and temperature. This transformation is not theatrical decoration; it materially alters the visual conditions under which the camera captures the fruit. The neural network’s perception shifts with the light. Shadows deepen, textures flatten, highlights bloom. The same object, under slightly different illumination, becomes statistically different to the model.

Classification here is not abstract computation. It is contextual and physical. It depends on photons, angles, proximity, and environmental variation. The viewer’s body participates in the production of the image that the algorithm evaluates. Seeing becomes relational.

Technically, the work was designed to hesitate. The model was trained on an intentionally overlapping dataset in which “maracuyá” and “passion fruit” are not cleanly separated. The images used for training include variations in ripeness, scale, color temperature, and framing. Rather than optimizing for maximal separation between classes, the training process preserves ambiguity at their boundary.

At runtime, the system does not simply output a label; it produces probability distributions. A dynamic confidence threshold determines when the model is “certain enough” to commit to a class. That threshold is not fixed. It is continuously modulated by live sensor input, including environmental light variation. Small contextual changes push the output probabilities back and forth across the decision boundary. The result is a controlled oscillation:
sometimes the system declares maracuyá,
sometimes passion fruit,
sometimes it withholds judgment altogether.

This flocking between categories is not an error. It is engineered ambiguity. Machine vision systems are often presented as instruments of certainty, extracting stable facts from unstable worlds. Threshold Navigator reverses that premise. It foregrounds the threshold itself — the invisible line where probability becomes assertion. By exposing and destabilizing that line, the work reveals classification as a negotiated event rather than an objective truth.

The fruit remains still. The light shifts. The viewer approaches. The model hesitates. In this system, recognition is not a final answer. It is a contingent agreement between matter, environment, data, and code. Classification becomes what it has always been:
a situated act.

Team

Leaders

Hugo Idarraga